Semantic-Aware Domain Generalized Segmentation Duo Peng 1 Yinjie Lei 1, Munawar Hayat 2 Yulan Guo 3 Wen Li 4 1 Sichuan University 2 Monash University

Size: px
Start display at page:

Download "Semantic-Aware Domain Generalized Segmentation Duo Peng 1 Yinjie Lei 1, Munawar Hayat 2 Yulan Guo 3 Wen Li 4 1 Sichuan University 2 Monash University"

Transcription

1 Semantic-Aware Domain Generalized Segmentation Duo Peng 1 Yinjie Lei 1, Munawar Hayat 2 Yulan Guo 3 Wen Li 4 1 Sichuan University 2 Monash University 3 Sun Yat-sen University 4 University of Electronic Science and Technology of China arxiv: v1 [cs.cv] 2 Apr 2022 duo peng@stu.scu.edu.cn, yinjie@scu.edu.cn, munawar.hayat@monash.edu guoyulan@sysu.edu.cn, liwenbnu@gmail.com Figure 1. Illustration of existing Instance Normalization and Whitening methods and our proposed approach. (a-c) Existing methods broadly eliminate the global distribution variance but ignore the category-level semantic consistency resulting in limited feature discrimination. (d) Our proposed modules (SAN & SAW) encourage both intra-category compactness and inter-category separation through category-level feature alignment leading to both effective style elimination and powerful feature discrimination. Abstract Deep models trained on source domain lack generalization when evaluated on unseen target domains with different data distributions. The problem becomes even more pronounced when we have no access to target domain samples for adaptation. In this paper, we address domain generalized semantic segmentation, where a segmentation model is trained to be domain-invariant without using any target domain data. Existing approaches to tackle this problem standardize data into a unified distribution. We argue that while such a standardization promotes global normalization, the resulting features are not discriminative enough to get clear segmentation boundaries. To enhance separation between categories while simultaneously promoting domain invariance, we propose a framework including two Corresponding Author: Yinjie Lei (yinjie@scu.edu.cn) novel modules: Semantic-Aware Normalization (SAN) and Semantic-Aware Whitening (SAW). Specifically, SAN focuses on category-level center alignment between features from different image styles, while SAW enforces distributed alignment for the already center-aligned features. With the help of SAN and SAW, we encourage both intra-category compactness and inter-category separability. We validate our approach through extensive experiments on widely-used datasets (i.e. GTAV, SYNTHIA, Cityscapes, Mapillary and BDDS). Our approach shows significant improvements over existing state-of-the-art on various backbone networks. Code is available at 1. Introduction Semantic segmentation is a critical machine vision task with multiple downstream applications, such as robotic nav-

2 igation [25, 36, 41, 64], autonomous vehicles [17, 27, 50, 63] and scene parsing [28, 68, 69, 71]. While the current fully supervised deep learning based segmentation methods can achieve promising results when they are trained and evaluated on data from same domains [1, 4 6, 21, 37, 40, 67, 70], their performance dramatically degrades when they are evaluated on unseen out-of-domain data. To enable generalization of models across domains, different domain adaptation techniques have been recently proposed [3, 15, 16, 23, 44, 46, 55, 62, 75]. However, a critical limitation of domain adaptation methods is their reliance on the availability of target domain in advance for training purposes. This is impractical for many real-world applications, where it is hard to acquire data for rarely occurring concepts. In this paper, we consider the challenging case of Domain Generalized Semantic Segmentation (DGSS), where we do not have access to any target domain data at the training time [9, 47, 48, 51, 66]. Existing methods tackle DGSS using two main approaches: (1) Domain Randomization [51, 66] which aims to increase the variety of training data by augmenting the source images to multiple domain styles. However, this is limiting since the augmentation schemes used are unable to cover different scenarios that may occur in the target domain. (2) Normalization and Whitening [9, 47, 48] which utilizes predefined Instance Normalization (IN) [61] or Instance Whitening (IW) [32] to standardize the feature distribution of different samples. IN separately standardizes features across each channel of individual images to alleviate the feature mismatch caused by style variations. However, as shown in Fig. 1 (a), IN only achieves center-level alignment and ignores the joint distribution among different channels. IW can remove linear correlation between channels, leading to well-clustered features of uniform distributions (see Fig. 1 (b)). Recent studies [9, 48] propose to combine IN and IW to achieve joint distributed feature alignment (see Fig. 1 (c)). Nevertheless, such global alignment strategy lacks the consideration of local feature distribution consistency. The features belonging to different object categories, which are originally well separated, are mapped together after normalization, leading to confusion among categories especially when generalizing to unseen target domains. Such semantic inconsistency inevitably results in sub-optimal training, causing performance degradation on unseen target domain and even the training domain (i.e. source domain). To address the inherent limitations of IN and IW, we propose two modules, Semantic-Aware Normalization (SAN) and Semantic-Aware Whitening (SAW), which collaboratively align category-level distributions aiming to enhance the discriminative strength of features (see Fig. 1 (d)). Compared with traditional IN&IW based methods, our approach brings two appealing benefits: First, it carefully integrates semantic-aware center alignment and distributed alignment, enabling both discriminative and compact matching of features from different styles. Therefore, our method can significantly enhance models generalization to out-of-domain distributed data. Second, existing methods improve the generalization ability at the cost of source domain performance [9, 47]. Nevertheless, our approach enhances the semantic consistency while improving category-level discrimination, thus leading to effective generalization with negligible performance drop on source domain. Our extensive empirical evaluations on benchmark datasets show that our approach improves upon previous DGSS methods, setting new state-of-the-art performances. Remarkably, our method also performs favorably compared with existing SOTA domain adaptation methods that are trained using target domain data. In summary, followings are the major contributions of our work. We propose effective feature alignment strategies to tackle out-of-domain generalization for segmentation, without access to target domain data for training. The proposed semantic-aware alignment modules, SAN and SAW, are plug-and-play and can easily be integrated with different backbone architectures, consistently improving their generalization performance. Through extensive empirical evaluations, and careful ablation analysis, we show the efficacy of our approach across different domains and backbones, where it significantly outperforms the current state-of-the-art. Remarkably, we even perform at par with approaches using target domain data for training purposes. 2. Background Here, we discuss recent approaches developed for Domain Adaptation (DA) and Domain Generalization (DG) in the context of semantic segmentation Domain Adaptation (DA) Domain Adaptation seeks to narrow the domain gap between the source and target domain data. It aims to enhance the generalization ability of the model by aligning the feature distributions between the source and target images [15, 16, 38, 39, 58 60]. Domain adaptation for semantic segmenation (DASS) was first studied in [24, 72], and since then has gained significant research attention. We can broadly categorize the existing approaches for DASS into Adversarial Training, and Self-Training based methods. Most of the existing work on DASS has been dominated by Adversarial Training based approaches [7, 15, 23, 56, 73]. Inspired by Generative Adversarial Networks [20], these approaches are generative in nature, and synthesize indistinguishable features which are domain-invariant and deceive the domain classifier. Self-Training based DASS approaches are relevant once labeled training data is scarce.

3 These methods [35, 75] train the model with pseudo-labels which are generated from the previous models predictions. However, DA methods require access to the samples from the target domain, which limits their applicability on totally unseen target domain Domain Generalization (DG) In contrast to Domain Adaptation, where the images in the target domain, although without labels, are accessible during the training process, Domain Generalization is evaluated on data from totally unseen domains [14, 43]. Domain generalization has been mostly explored on the image classification task, and a number of approaches have been proposed using as meta-learning [2, 29, 30, 34], adversarial training [31, 33, 52], autoencoders [18, 33], metric learning [12, 42] and data augmentation [19, 74]. The research on domain generalization for semantic segmentation (DGSS) is still in its infancy, with only a few existing approaches [9, 47, 48, 51, 66]. These existing DGSS methods mainly focus on two aspects: (1) Domain Randomization and (2) Normalization and Whitening. Domain Randomization based methods seek to synthesize images with different styles e.g. [66] leverages the advanced image-to-image translation to transfer a source domain image to multiple styles aiming to learn a model with high generalizability. Similarly, GTR [51] randomizes the synthetic images with the styles of unreal paintings in order to learn domain-invariant representations. Normalization and Whitening Methods apply different normalization techniques such as Instance Normalization (IN) [61] or whitening [48]. For example, based on the observation that Instance Normalization (IN) [61] prevents overfitting on domain-specific style of training data, [47] proposes to utilize IN to capture style-invariant information from appearance changes while preserving content related information. Inspired from [47], [47] proposes Switchable Whitening (SW), which combines IN with other whitening methods, aiming to achieve a flexible and generic features. In another recent approach [9], an instance selective whitening to disentangle domain-specific and domain-invariant properties is explored and only domain-specific features are normalized and whitened. However, all aforementioned methods perform a global alignment for features belong to different image categories. We aim to address this crucial limitation and propose an approach which enforces local semantic consistency during the trend of global style elimination. 3. Preliminaries Let s denote an intermediate mini-batch feature map by F R N K H W, where N, K, H and W are the dimensions of the feature map, i.e. batch sample, channel, height and width, respectively. F n,k,h,w F represent the feature element, where n, k, h, w respectively indicate the index of each dimension. Similarly, F n R K H W denotes the features of n-th sample from mini-batch, and F n,k R H W denotes the k-th channel ofn-th sample. Below, we first define Instance Normalization (IN) and Instance Whitening (IW), which have been commonly used by the existing approaches. Instance Normalization (IN) simply standardizes features using statistics (i.e. mean and standard deviation) computed over each individual channel from each individual sample, given by: IN(F) = F n,k µ n,k σ n,k +ε, (1) where IN( ) denotes the instance normalization process and ε is a small value to avoid division by zero. The meanµ n,k and standard deviationσ n,k ofn-th samplek-th channel are computed as follows: µ n,k = 1 HW H h=1w=1 h=1w=1 W F n,k,h,w, (2) σ n,k = 1 H W (F n,k,h,w µ n,k ) HW 2. (3) Using above operations, IN transforms the features from different image samples to have a standard distribution, i.e, zero mean and one standard deviation. However, even though the features of each channel are centered and scaled into standard distribution, the joint distribution between channels might be mismatched. Instance Whitening (IW) standardizes features by decorrelating the channels. As shown in Fig. 3 (a), it removes correlation between channels by making the covariance matrix close to the identity matrix through the following objective function: N L IW = Ψ(F n ) I 1, (4) n=1 where Ψ( ) and I denotes the channel correlation and identity matrix. Ψ(F n ) is defined as: Cov(F n,1,f n,1 ) Cov(F n,1,f n,k ) Ψ(F n) = , (5) Cov(F n,k,f n,1 ) Cov(F n,k,f n,k ) wherecov(f n,i,f n,j ) is the covariance value between the i-th channel andj-th channel of featuref n, given by: Cov(F n,i,f n,j ) = 1 HW H h=1w=1 W (F n,i,h,w µ n,i )(F n,j,h,w µ n,j ). IW [32] is capable of unifying the joint distribution shape through channel decorrelation for each sample. Combined with IN [61], IW can make a unified joint distributed (6)

4 NKHW NCHW C C C g C Y NKHW obj obj c c c k c Figure 2. The detailed architecture of our Semantic Aware Normalization (SAN) module. SAN adapts a multi-branch normalization strategy, aiming to transform the feature mapfinto the category-level normalized features F, that are semantic-aware center aligned. alignment. However, such global matching might cause some features to be mapped to an incorrect semantic category, resulting in poor segmentation boundary decisions. In the following Sec. 4, we introduce our method which aims to tackle these problems, and preserve the semantic relationships between different categories. 4. Proposed Method Our goal is to achieve semantic-aware center alignment and distributed alignment. For this, we introduce two novel modules, Semantic-Aware Normalization (SAN) and Semantic-Aware Whitening (SAW). We sequentially embed these two modules in our network as shown in Fig. 1. We discuss these modules in detail below Semantic Aware Normalization (SAN) Given an intermediate mini-batch feature map F, SAN transforms F into a feature map that is category-level centred. With the help of segmentation labelsy, we can easily obtain the desired objective feature mapf obj as: F obj = Fc n,k µc n,k σ c n,k +ε γ c +β c, (7) µ c n,k = 1 F c Y(c) n,k, (8) Y(c) σ c n,k = 1 Y(c) (F c n,k µc n,k )2, (9) Y(c) where µ c n,k and σc n,k are the mean and standard deviation computed fromc-th category features of k-th channel,n-th sample, and thec-th category labely(c). FeaturesF c n,k belong to c-th category in channelf nk. The weights for scaling and shifting are denoted byγ andβ, respectively, which are both learnable parameters. We seperately allocate these affine parameters for each category (i.e. γ c and β c ) aiming to adjust the standardized features from different categories to distinct spaces, thus making our feature space more discriminative. Note that different samples in the mini-batch share the same affine parameters in order to cast features of same category into a same feature space, thus ensuring category-level center alignment. We utilize SAN to approximate Eq. (7) since the label Y is unavailable when testing on target domains. We follow a series of steps to transform the input features F into the objective features F obj as shown in Fig. 2. First, we leverage the segmentation masks generated from the classifier to highlight the category region while simultaneously suppressing other regions in each normalization branch. F c = F M c, (10) wherem c denotes the mask of thec-th category, denotes Hadamard product and F c represents the masked feature map in c-th category branch. The generated mask can be too rough to precisely locate category features. We therefore introduce a Category-level Feature Refinement (CFR) block to adaptively adjust the highlighted features F c into, given by: F c F c = Sigm(f 3 3 ([F c,max;f c,avg;m c ])) F c, (11) where f 3 3 ( ) and Sigm( ) denote 3 3 convolution and sigmoid function, respectively. F c,max and F c,avg are max-pooled and average-pooled features of feature mapf c. [a; b] is the concatenation of a and b along channel axis. In order to further refine the category-level center alignment, we design a Regional Normalization layer which normalizes features only within the category region instead of whole scene. After refinement, only the feature elements

5 K K K K C K K w > w >... > w w > w >... > w w C > w C >... > w C M K M K M K M w w... w K w w... w K w C w C... w K C K M K M K w w w C w w w C K/C C C C C C C C C C C Figure 3. Illustration of feature whitening in IW [32], GIW [8] and the proposed SAW. (a) IW de-correlates all channels from each other. (b) GIW only de-correlates the channels in the same group. (c) SAW allocates channels related to different categories in each group. with high value are assigned the category region. To flexibly identify feature elements, we apply k-means clustering on the spatial feature map obtained by avgpooling along channel axis. After dividing the spatial elements into k clusters, the clusters from the first to thet-th are considered to be the category region, and the remaining clusters are considered as ignored region. We set t to one and search the optimal k through the hyper-parameter search. In this paper, k is set to 5. See Sec. 5.5 for more details. Thus, we can assign the feature elements of c-th branch into 2 clusters{φ c low,φc high }, whereφc high denotes the identified category region. We normalize features within the category regionφ c high for each individual channel. Finally, all category branches are added together, then re-shifted and scaled by the learnable affine parameters: F = C c=1 RN(F c,φ c high) γ c +β c, (12) where RN(,Φ c high ) denotes our regional normalization in c-th branch, γ c and β c are affine parameters as per Eq. (7). In order to ensure the processed feature map F are categorylevel-center-aligned as per Eq. (7), we optimize the following cross-entropy loss: L SAN = CE(M,Y)+ F F obj 1, (13) where M {M 1,M 2,...,M C } denotes the set of predicted segmentation masks Semantic Aware Whitening (SAW) We propose the Semantic-Aware Whitening (SAW) module, to further enhance channel decorrelation for the distributed alignment of the already semantic-centred features. Instance Whitening (IW) is capable of unifying the joint distribution, which is useful for distributed alignment. However, directly adopting IW is not feasible, since such strong whitening that strictly removes correlation between all channels may damage the semantic content, resulting in loss of crucial domain-invariant information. Group Instance Whitening (GIW [8], shown in Fig. 3 (b)) is a simple solution to this problem. Given the feature map F which is the output of SAN module, GIW is defined by: G m n = [ F n, K(m 1) M L GIW = 1 N +1; F n, K(m 1) M N n=1m=1 +2;...; F n, Km], (14) M M Ψ(G m n ) I 1, (15) where G m n denotes the m-th group of n-th sample, M is the number of groups, and Ψ( ) is the channel correlation matrix defined in Eq. (4). While GIW improves the generalization by partial (group) channel decorrelation, it strictly decorrelates neighboring channels, lacking the consideration of searching more appropriate channel combinations. It is well-known that the channels in convolution neural networks are highly related to semantics. To this end, we propose SAW module to rearrange channels, ensuring each group contains channels related to different categories. Compared with grouping same-category-related channels, decorrelation between channels from different categories is more reasonable, since different-category-related channels activate different regions. Removing the correlations between those channels not only enhances the representation

6 capacity of each single channel, but also prevents the information loss, resulting in distributed alignment with minor changes to the semantic content. As for segmentation model, the segmentation results of each category are obtained by multiplying all the channels by their corresponding weights and then adding them up. The weight value determines the influence of channel on the category. Hence, to identify each channel belongs to which category, we utilize the classifier from the SAN module. As shown in Fig. 3 (c), for each category, there are K weights corresponding to K channels, i.e. {wclass 1 c,w2 class c,...,wk class c } where c {1,...,C}. After turning them into absolute values. We rank the weights of each category from the largest to the smallest. Then in each category, the first K C weight indexes are selected. For the sake of clarity, we use I R C K C to denote the all selected indexes, where I(i, j) represents the j-th selected index of i-th category, i {1,...,C} and j {1,..., K C }. We arrange K C groups and allocate C different-categoryrelated channels for each group. Specifically, each channel is weighted by its corresponding classifier weight before grouping, aiming to execute adaptive whitening transformation. Therefore, the m-th group of n-th sample: G m n in Eq. (14) can be modified asḡ m n : Ḡ m n = [ F n,i(1,m) w I(1,m) class1 ; F n,i(2,m) w I(2,m) class2 ;...; F n,i(c,m) w I(C,m) classc ]. (16) Correspondingly, our whitening loss is formulated as: L SAW = 1 N N K C n=1m=1 Ψ(Ḡ m n ) I 1. (17) Note that operations of SAW module do not change the features of main network in forward pass. Therefore, different from SAN module, SAW is only applied during training. 5. Experiments 5.1. Datasets Description Synthetic Datasets. GTA5 [53] is a synthetic image dataset, which is collected by using GTA-V game engine. It contains images with a resolution of along with their pixel-wise semantic labels. SYNTHIA [54] is a large synthetic dataset with pixel-level semantic annotations. The subset SYNTHIA-RANDCITYSCAPES [54] is used in our experiments which contains 9400 images with a high resolution of Real-World Datasets. Cityscapes [10] is a high resolution dataset (i.e ) of 5000 vehicle-captured urban street images taken from from 50 different cities primarily in Germany. BDDS [65] contains thousands of realworld dashcam video frames with accurate pixel-wise annotations, where images are provided with a resolution of Mapillary [45] contains images with diverse resolutions. The annotations contain 66 object categories, but only 19 categories overlap with others Implementation Details We initialize the weights of the feature extractor module with an ImageNet [11] pre-trained model. We use SGD [26] optimizer with with an initial learning rate of 5e-4, a batch size of 2, a momentum of 0.9 and a weight decay of 5e-4. Besides, we follow the polynomial learning rate scheduling [6] with the power of 0.9. We train model for iterations. All datasets have 19 common categories, thus the parameter C defined in both SAN and SAW is set to 19. However, since SAW arranges K C groups, C can only be 2, 4, 8 or 16 to make K (channel number) divisible by C. Based on the results of ablation study on parameter C (Sec. 5.5 ), we set C = 4 in both SAN and SAW. We implement our method on PyTorch [49] and use a single NVIDIA RTX 3090 GPU for our experiments. Following previous works, we use PASCAL VOC Intersection over Union (IoU) [13] as the evaluation metric Comparison with DG and DA methods For brevity, we use G, S, C, B and M to denote GTA5 [53], SYNTHIA [54], Cityscapes [10], BDDS [65] and Mapillary [45], respectively. We extensively evaluate our method with different backbones including VGG-16 [57], ResNet-50 and ResNet-101 [22]. We repeat each experiment three times, and report the average results. Unlike existing synthetic-to-real generalization approaches, we propose to evaluate our model from an arbitrary domain to other unseen domains. Therefore, we conduct comprehensive experiments on five generalization settings, from (1) G to C, B, M & S; (2) S to C, B, M & G; (3) C to G, S, B & M; (4) B to G, S, C & M; (5) M to G, S, C & B. Our results reported in Tab. 1 on the first three settings, and Appendix A on the remaining 2 settings, suggest that our model consistently gains the best performance across all settings and backbones. Compared to the the second best results (see underlined values), our method shows a large improvement. Visual samples for qualitative comparison are given in Fig. 4. Remarkably, our method performs favorably in comparison to the methods that have access to the target domain data, see results in Appendix B Source Domain Performance Decay Analysis A common pitfall of the domain generalization methods is that their performance degrades on the source domain. To compare different methods for this aspect, we evaluate them on the test set of the source domain in Tab. 2. The results suggest that our method largely retains performance on the source domain, and performs comparably with the model trained without domain-generalization (Baseline in Tab. 2).

7 Table 1. Performance comparison in terms of miou (%) between Domain Generalization methods. The best and second best results are highlighted and underlined, respectively. denotes our re-implemention of the respective method. G, C, B, M and S denote GTA5 [53], Cityscapes [10], BDDS [65], Mapillary [45] and SYNTHIA [54], respectively. Methods Publication Backbone Train on GTA5 (G) Train on SYNTHIA (S) Train on Cityscapes (C) C B M S C B M G B M G S Baseline IBN [47] ECCV SW [48] ICCV DRPC [66] ICCV 2019 VGG GTR [51] TIP ISW [9] CVPR Ours Baseline IBN [47] ECCV SW [48] ICCV DRPC [66] ICCV 2019 ResNet GTR [51] TIP ISW [9] CVPR Ours Baseline IBN [47] ECCV SW [48] ICCV DRPC [66] ICCV 2019 ResNet GTR [51] TIP ISW [9] CVPR Ours Figure 4. Visual comparison with different Domain Generalization methods on unseen domains i.e. Cityscapes [10], BDDS [65], Mapillary [45] and SYNTHIA [54], with the model trained on GTA5 [53]. The backbone network is ResNet Ablation Study SAN and SAW. We investigate the individual contribution of SAN and SAW modules towards overall performance. Tab. 3 shows the miou improvement on ResNet- 50 once we progressively integrate SAN and SAW. Experiments are conducted for generalization from GTA5 (G) to the other four datasets i.e. Cityscapes (C), BDDS (B), Mapillary (M) and SYNNTHIA(S). In our case, each of them helps boost the generalization performance by a large margin. Specifically, we observe that SAN and SAW greatly achieve an average improvement of 8.71% and 7.93%, respectively. We further observe that the models with only SAW show slightly weaker generalization capacity than those with only SAN. This is because without categorylevel centering of SAN, SAW performs distributed alignment, which may lead to incorrect feature matching between different categories. Therefore, our SAW module is mainly proposed to complement SAN in an integrated approach. As shown in Tab. 3, the best performance is achieved with a combination of both SAN and SAW. CFR block in SAN. To demonstrate the effectiveness of the Category-level Feature Refinement (CFR) in SAN, we conducted an ablation experiment by removing the CFR block. As shown in Tab. 4 (top row), model without CFR consistently performs worse than model with SAN. With the help of CFR, SAN achieves an average gain of 2.30%, which demonstrates its usefulness. Category-related Grouping in SAW. We perform ablations on different grouping strategies to verify their contributions. We conduct experiments on baseline with IW, GIW and SAW, respectively. The architectures of these

8 Table 2. Comparison of different methods for performance drop on source domain. The performance drop ( ) is obtained with respect to the baseline. The best and second best values are highlighted and underlined, respectively. The network backbone is ResNet-50. Methods GTA5 SYNTHIA Cityscapes BDDS Mapillary Baseline C IBN [47] SW [48] DRPC [66] GTR [51] ISW [9] Ours Table 3. Ablation analysis of SAN (Sec. 4.1) and SAW (Sec. 4.2). k Figure 5. Change in performance with hyper-parameters: C and k. The experimental backbone is ResNet-50. three models are illustrated in Fig. 3. As shown in Tab. 4 (bottom row), by adopting the general grouping operation, GIW shows significant improvement compared to model with IW. When applying our proposed SAW which performs category-related grouping, the performance of network improves to 37.54%, 34.97%, 39.85% & 28.46% on G C, B, M & S. This confirms the effectiveness of categoryrelated grouping in SAW. Hyper-Parameter Analysis. C decides on the number of categories for semantic-aware feature alignment in both SAN and SAW. Since we rank the categories according to the their respective proportions in training data, only the first C categories are selected to handle category-level feature matching. To ensure that the channel dimension of feature maps is divisible by C, we can only set C {2,4,6,8,16}. Fig. 5 (a) shows compares results for different values, suggesting the optimalc is 4. In CFR block, we adopt k-means clustering to separate feature elements into category region and background. For searching the optimal parameter k, we only choose the first cluster whose cluster center is highest as the category region. As shown in Fig. 5 (b), model performs best when adopting k=5. 6. Conclusion and limitations In this manuscript, we present a domain generalization approach to address the out-of-domain generalization for semantic segmentation. We propose two novel modules: Semantic-Aware Normalization (SAN) and Semantic- Aware Whitening (SAW), which sequentially perform category-level center alignment and distributed alignment to achieve both domain-invariant and discriminative features. Comprehensive experiments demonstrate the effectiveness of SAN and SAW with state-of-the-art perfor- Methods SAN SAW Train on GTA5 (G) C B M S Baseline SAN SAW All Table 4. Ablation of different blocks in SAN and SAW Methods Train on GTA5 (G) C B M S Baseline Baseline + SAN (w/o CFR) Baseline + SAN Baseline + IW Baseline + GIW Baseline + SAW mance in both domain generalization and domain adaptation. Although the method effectively eliminates feature differences caused by style variations on source domain, the extracted style-invariant features may still contain sourcedomain specific cues, leading to significant performance delta between the source and target domain. Some interesting future directions to close this gap include learning how to model domain shift (meta-learning), integrating multiple domain-specific neural networks (ensemble learning) and teaching a model to perceive generic features regardless of the target task (disentangled representation learning). The datasets used in the paper lack diversity and have biases as they are mostly captured in the developed world. Beyond that, this paper have no ethical problem including personally identifiable information, human subject experimentation and military application. Acknowledgement: This work was partially supported by the National Natural Science Foundation of China (No , U20A20185). M. Hayat is supported by ARC DECRA fellowship DE

9 References [1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 39(12): , [2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 31: , [3] Shai Ben-David, John Blitzer, Koby Crammer, Fernando Pereira, et al. Analysis of representations for domain adaptation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 19:137, [4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In Proceedings of the International Conference on Learning Representations (ICLR), pages , [5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 40(4): , [6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arxiv preprint arxiv: , , 6 [7] Yuhua Chen, Wen Li, and Luc Van Gool. Road: Reality oriented adaptation for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [8] Wonwoong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [9] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , , 3, 7, 8 [10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , , 7 [11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [12] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 32: , [13] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision (IJCV), 111(1):98 136, [14] Chuang Gan, Tianbao Yang, and Boqing Gong. Learning attributes equals multi-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 87 97, [15] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the International Conference on Machine Learning (ICML), pages , [16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. the Journal of Machine Learning Research (JMLR), 17(1): , [17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [18] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [19] Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. Dlow: Domain flow for adaptation and generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages , [20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11): , [21] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), pages , [24] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcns in the wild: Pixel-level adversarial and constraint-based adaptation. arxiv preprint arxiv: ,

10 [25] Wonsuk Kim and Junhee Seok. Indoor semantic segmentation for robot navigating on mobile. In Proceedings of the IEEE International Conference on Ubiquitous and Future Networks (ICUFN), pages 22 25, [26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 25: , [27] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 61 71, [28] Yinjie Lei, Duo Peng, Pingping Zhang, Qiuhong Ke, and Haifeng Li. Hierarchical paired channel fusion network for street scene change detection. IEEE Transactions on Image Processing, 30:55 67, [29] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), [30] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [31] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [32] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. arxiv preprint arxiv: , , 3, 5 [33] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages , [34] Yiying Li, Yongxin Yang, Wei Zhou, and Timothy Hospedales. Feature-critic networks for heterogeneous domain generalization. In Proceedings of the International Conference on Machine Learning (ICML), pages , [35] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [36] Hao Liu, Yulan Guo, Yanni Ma, Yinjie Lei, and Gongjian Wen. Semantic context encoding for accurate 3d point cloud segmentation. IEEE Transactions on Multimedia, 23: , [37] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [38] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the International Conference on Machine Learning (ICML), pages , [39] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. arxiv preprint arxiv: , [40] Yanni Ma, Yulan Guo, Hao Liu, Yinjie Lei, and Gongjian Wen. Global context reasoning for semantic segmentation of 3d point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages , [41] Ryusuke Miyamoto, Yuta Nakamura, Miho Adachi, Takeshi Nakajima, Hiroki Ishida, Kazuya Kojima, Risako Aoki, Takuro Oki, and Shingo Kobayashi. Vision-based roadfollowing using results of semantic segmentation for autonomous navigation. In Proceedings of the IEEE International Conference on Consumer Electronics (ICCE), pages , [42] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [43] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In Proceedings of the International Conference on Machine Learning (ICML), pages 10 18, [44] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [45] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , , 7 [46] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [47] Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages , , 3, 7, 8 [48] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaoou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , , 3, 7, 8 [49] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban

11 Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In Proceedings of the Advances in Neural Information Processing Systems Workshops (NuerIPS Workshops), [50] Duo Peng, Yinjie Lei, Wen Li, Pingping Zhang, and Yulan Guo. Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages , [51] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing (TIP), 30: , , 3, 7, 8 [52] Mohammad Mahfujur Rahman, Clinton Fookes, Mahsa Baktashmotlagh, and Sridha Sridharan. Correlation-aware adversarial domain adaptation and generalization. Pattern Recognition (PR), 100:107124, [53] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Proceedings of the European Conference on Computer Vision (ECCV), pages , , 7 [54] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , , 7 [55] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [56] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arxiv preprint arxiv: , [58] Kihyuk Sohn, Sifei Liu, Guangyu Zhong, Xiang Yu, Ming- Hsuan Yang, and Manmohan Chandraker. Unsupervised domain adaptation for face recognition in unlabeled videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [59] Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [60] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [61] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , , 3 [62] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [63] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [64] Xiaoqing Ye, Jiamao Li, Hexiao Huang, Liang Du, and Xiaolin Zhang. 3d recurrent neural networks with context fusion for point cloud semantic segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages , [65] Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving video database with scalable annotation tooling. arxiv preprint arxiv: , 2(5):6, , 7 [66] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , , 3, 7, 8 [67] Pingping Zhang, Wei Liu, Yinjie Lei, and Huchuan Lu. Semantic scene labeling via deep nested level set. IEEE Transactions on Intelligent Transportation Systems, 22(11): , [68] Pingping Zhang, Wei Liu, Yinjie Lei, Hongyu Wang, and Huchuan Lu. Deep multiphase level set for scene parsing. IEEE Transactions on Image Processing, 29: , [69] Pingping Zhang, Wei Liu, Yinjie Lei, Hongyu Wang, and Huchuan Lu. Rapnet: Residual atrous pyramid network for importance-aware street scene parsing. IEEE Transactions on Image Processing, 29: , [70] Pingping Zhang, Wei Liu, Hongyu Wang, Yinjie Lei, and Huchuan Lu. Deep gated attention networks for largescale street-level scene segmentation. Pattern Recognition, 88: , [71] Rui Zhang, Sheng Tang, Yongdong Zhang, Jintao Li, and Shuicheng Yan. Scale-adaptive convolutions for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [72] Yang Zhang, Philip David, and Boqing Gong. Curriculum domain adaptation for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages ,

12 [73] Yiheng Zhang, Zhaofan Qiu, Ting Yao, Dong Liu, and Tao Mei. Fully convolutional adaptation networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [74] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages , [75] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages , , 3

13 arxiv: v1 [cs.cv] 2 Apr 2022 Appendix A. Evaluation on other DG settings Existing DG methods [3, 11 13, 20] only focus on three domain generalization settings, i.e. (1) G C, B, M, & S; (2) S C, B, M, & G and (3) C G, S, B, & M while losing the sight of the other two DG settings: (4) B G, S, C, & M and (5) M G, S, C, & M. However, recently several studies [2, 6] stress the importance of the last two settings. Therefore, we consider a more comprehensive evaluation which performs generalization from each of them. Results on the last two settings are reported in Tab. 1, suggesting that our model consistently achieves the state-of-the-art results on all settings and backbones. Table 1. Performance comparison in terms of miou (%) between DG methods. The best and second best results are highlighted and underlined, respectively. denotes our re-implemention of the respective method. G, C, B, M and S denote GTA5, Cityscapes, BDDS, Mapillary and SYNTHIA, respectively. Methods Backbone Train on BDDS (B) Train on Mapillary (M) G S C M G S C B Baseline IBN [11] SW [12] DRPC [20] VGG GTR [13] ISW [3] Ours Baseline IBN [11] SW [12] DRPC [20] ResNet GTR [13] ISW [3] Ours Baseline IBN [11] SW [12] DRPC [20] ResNet GTR [13] ISW [3] Ours Appendix B. Comparison with DA methods Domain Adaptation (DA) methods require access to the target domain to solve domain shift problems. In contrast, our method is designed in Domain Generalization (DG) manner for broad generalization to totally unseen domains without accessing any target domain data. Therefore, the target domain-accessible DA methods have the inherent performance superiority than DG methods which are target domain-agnostic. In order to see whether our approach is up to the performance standard of DA, we compare the results of our method with those reported from several previous state-of-the-art DA methods. From Tab. 2 and Tab. 3, we can see that the generalization performance of our method outperforms the adaptation performance of most other techniques. In addition, no target-domain data is needed in our method, resulting in more extensive applicability. Table 2. Comparison results between ours and Domain Adaptation methods on GTA5 Cityscapes. DA and DG denote Domain Adaption and Domain Generalization respectively. Backbone Task Method Access Tgt miou VGG-16 Resnet-101 FCN wild [6] 27.1 CDA [21] 28.9 CyCADA [5] 34.8 ROAD [1] 35.9 I2I [9] 31.8 AdaptSegNet [14] 35.0 DA SSF-DAN [4] 37.7 DCAN [18] 36.2 CBST [22] 30.9 CLAN [8] 36.6 ADVENT [16] 36.1 DPR [15] 37.5 BDL [7] 41.3 FDA [19] 42.2 DG Ours 38.2 CyCADA [5] 42.7 ROAD [1] 39.4 I2I [9] 35.4 AdaptSegNet [14] 41.4 DA DCAN [18] 41.7 CLAN [8] 43.2 ADVENT [16] 43.8 DPR [15] 46.5 IntraDA [10] 46.3 DADA [17] 47.3 DG Ours 45.3 Table 3. Comparison results between ours and Domain Adaptation methods on SYNTHIA Cityscapes. Backbone Task Method Access Tgt miou VGG-16 Resnet-101 FCN wild [6] 20.2 CDA [21] 29.0 ROAD [1] 36.2 DCAN [18] 35.4 DA CBST [22] 35.4 ADVENT [16] 31.4 DPR [15] 33.7 BDL [7] 39.0 FDA [19] 40.5 DG Ours 37.4 ADVENT [16] 40.8 DA DPR [15] 40.0 IntraDA [10] 41.7 DADA [17] 42.6 DG Ours 40.9 Appendix C. Further Implementation Details We follow previous work [3, 11] to adopt normalization and whitening at the first two stages of convolution layers, since shallow layers encode more style information [11]. As shown in Fig. 1, for each backbone network, we impose SAN and SAW after stage 1 and stage 2.

14 H W H/ W/ H/ W/ H/ W/ H/W/ H/W/ H W C H W H/ W/ H/ W/ H/ W/ H/W/ H/W/ H W C Figure 1. Detailed Architecture of our approach with the backbone of VGG and ResNet. Appendix D. Computational complexity As shown in Tab. 4, compared to the baseline, our methods performs domain generalization with negligible addition in both training and inference time. This is because the proposed modules are only implemented in the first two layers of the network, for only four main categories i.e. C = 4, see Sec. 5.5 of the main paper. The additional memory overhead from our modules is less than 2G. Table 4. Comparison on computation cost. Backbone Methods Memory (G) Training Time (s) Inference Time (ms) Vgg-16 Res-50 Res-101 Baseline Ours Baseline Ours Baseline Ours Appendix E. More qualitative results Fig. 2 shows more qualitative results under various unseen domains. We demonstrate the effects of the proposed semantic-aware feature matching by comparing the segmentation results from our proposed approach and the baseline. In the setting of GTA5 Mapillary, the baseline fails to cope with these weather changes, while ours still shows fair results. Under the illumination changes as shown in GTA5 BDDS, our method finds the road and sidewalk clearer than the baseline. References [1] Yuhua Chen, Wen Li, and Luc Van Gool. Road: Reality oriented adaptation for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [2] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [3] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [4] Liang Du, Jingang Tan, Hongye Yang, Jianfeng Feng, Xiangyang Xue, Qibao Zheng, Xiaoqing Ye, and Xiaolin Zhang. Ssf-dan: Separated semantic feature based domain adaptation network for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [5] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), pages , [6] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcns in the wild: Pixel-level adversarial and constraint-based adaptation. arxiv preprint arxiv: , [7] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [8] Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Taking a closer look at domain shift: Category-level adversaries for semantics consistent domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [9] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [10] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages , [11] Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages ,

15 %DVHOLQH 2XUV *URXQG 7UXWK *7$ ė0dsloodu\ *7$ ė%''6 *7$ ė&lw\vfdshv *7$ ė6<17+,$ 8QVHHQ,PDJHV URDG VN\ VZDON SHUVRQ EXLOG ULGHU ZDOO FDU IHQFH WUXFN SROH EXV WOLJKW WUDLQ WVLJQ YHJ WHUUDLQ PF\FOH ELF\FOH XQODEHO Figure 2. Qualitative results of our approach generalizing from GTA5 to other four domains. [12] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaoou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages , [13] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing (TIP), 30: , [14] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker.

COCO18-DensePose-BUPT-PRIV

COCO18-DensePose-BUPT-PRIV * Beijing University of Posts and Telecommunications (BUPT) COCO 2018 DensePose Test AP AP50 AP75 APm APl BUPT-PRIV (Ours) 64 92 75 57 67 PlumSix 58 89 66 50 61 ML_Lab 57 89 64 51 59 Sound of silent 57

More information

國家圖書館典藏電子全文

國家圖書館典藏電子全文 i ii Abstract The most important task in human resource management is to encourage and help employees to develop their potential so that they can fully contribute to the organization s goals. The main

More information

國立中山大學學位論文典藏.PDF

國立中山大學學位論文典藏.PDF I II III The Study of Factors to the Failure or Success of Applying to Holding International Sport Games Abstract For years, holding international sport games has been Taiwan s goal and we are on the way

More information

附件1:

附件1: 附 件 1: 全 国 优 秀 教 育 硕 士 专 业 学 位 论 文 推 荐 表 单 位 名 称 : 西 南 大 学 论 文 题 目 填 表 日 期 :2014 年 4 月 30 日 数 学 小 组 合 作 学 习 的 课 堂 管 理 攻 硕 期 间 及 获 得 硕 士 学 位 后 一 年 内 获 得 与 硕 士 学 位 论 文 有 关 的 成 果 作 者 姓 名 论 文 答 辩 日 期 学 科 专

More information

报 告 1: 郑 斌 教 授, 美 国 俄 克 拉 荷 马 大 学 医 学 图 像 特 征 分 析 与 癌 症 风 险 评 估 方 法 摘 要 : 准 确 的 评 估 癌 症 近 期 发 病 风 险 和 预 后 或 者 治 疗 效 果 是 发 展 和 建 立 精 准 医 学 的 一 个 重 要 前

报 告 1: 郑 斌 教 授, 美 国 俄 克 拉 荷 马 大 学 医 学 图 像 特 征 分 析 与 癌 症 风 险 评 估 方 法 摘 要 : 准 确 的 评 估 癌 症 近 期 发 病 风 险 和 预 后 或 者 治 疗 效 果 是 发 展 和 建 立 精 准 医 学 的 一 个 重 要 前 东 北 大 学 中 荷 生 物 医 学 与 信 息 工 程 学 院 2016 年 度 生 物 医 学 与 信 息 工 程 论 坛 会 议 时 间 2016 年 6 月 8 日, 星 期 三,9:30 至 16:00 会 议 地 址 会 议 网 址 主 办 单 位 东 北 大 学 浑 南 校 区 沈 阳 市 浑 南 区 创 新 路 195 号 生 命 科 学 大 楼 B 座 619 报 告 厅 http://www.bmie.neu.edu.cn

More information

2/80 2

2/80 2 2/80 2 3/80 3 DSP2400 is a high performance Digital Signal Processor (DSP) designed and developed by author s laboratory. It is designed for multimedia and wireless application. To develop application

More information

Microsoft PowerPoint _代工實例-1

Microsoft PowerPoint _代工實例-1 4302 動態光散射儀 (Dynamic Light Scattering) 代工實例與結果解析 生醫暨非破壞性分析團隊 2016.10 updated Which Size to Measure? Diameter Many techniques make the useful and convenient assumption that every particle is a sphere. The

More information

PowerPoint Presentation

PowerPoint Presentation Decision analysis 量化決策分析方法專論 2011/5/26 1 Problem formulation- states of nature In the decision analysis, decision alternatives are referred to as chance events. The possible outcomes for a chance event

More information

acl2017_linguistically-regularized-lstm-MinlieHuang

acl2017_linguistically-regularized-lstm-MinlieHuang ACL 2017 Linguistically Regularized LSTM for Sentiment Classification Qiao Qian, Minlie Huang, Jinhao Lei, Xiaoyan Zhu Dept. of Computer Science Tsinghua University 1 aihuang@tsinghua.edu.cn Outline Introduction

More information

Improved Preimage Attacks on AES-like Hash Functions: Applications to Whirlpool and Grøstl

Improved Preimage Attacks on AES-like Hash Functions: Applications to Whirlpool and Grøstl SKLOIS (Pseudo) Preimage Attack on Reduced-Round Grøstl Hash Function and Others Shuang Wu, Dengguo Feng, Wenling Wu, Jian Guo, Le Dong, Jian Zou March 20, 2012 Institute. of Software, Chinese Academy

More information

Microsoft Word - 01李惠玲ok.doc

Microsoft Word - 01李惠玲ok.doc 康 寧 學 報 11:1-20(2009) 1 數 位 學 習 於 護 理 技 術 課 程 之 運 用 與 評 值 * 李 惠 玲 ** 高 清 華 *** 呂 莉 婷 摘 要 背 景 : 網 路 科 技 在 教 育 的 使 用 已 成 為 一 種 有 利 的 教 學 輔 助 工 具 網 路 教 學 的 特 性, 在 使 學 習 可 不 分 時 間 與 空 間 不 同 進 度 把 握 即 時 性 資

More information

08陈会广

08陈会广 第 34 卷 第 10 期 2012 年 10 月 2012,34(10):1871-1880 Resources Science Vol.34,No.10 Oct.,2012 文 章 编 号 :1007-7588(2012)10-1871-10 房 地 产 市 场 及 其 细 分 的 调 控 重 点 区 域 划 分 理 论 与 实 证 以 中 国 35 个 大 中 城 市 为 例 陈 会 广 1,

More information

UDC Empirical Researches on Pricing of Corporate Bonds with Macro Factors 厦门大学博硕士论文摘要库

UDC Empirical Researches on Pricing of Corporate Bonds with Macro Factors 厦门大学博硕士论文摘要库 10384 15620071151397 UDC Empirical Researches on Pricing of Corporate Bonds with Macro Factors 2010 4 Duffee 1999 AAA Vasicek RMSE RMSE Abstract In order to investigate whether adding macro factors

More information

10384 19020101152519 UDC Rayleigh Quasi-Rayleigh Method for computing eigenvalues of symmetric tensors 2 0 1 3 2 0 1 3 2 0 1 3 2013 , 1. 2. [4], [27].,. [6] E- ; [7], Z-. [15]. Ramara G. kolda [1, 2],

More information

Shanghai International Studies University THE STUDY AND PRACTICE OF SITUATIONAL LANGUAGE TEACHING OF ADVERB AT BEGINNING AND INTERMEDIATE LEVEL A Thes

Shanghai International Studies University THE STUDY AND PRACTICE OF SITUATIONAL LANGUAGE TEACHING OF ADVERB AT BEGINNING AND INTERMEDIATE LEVEL A Thes 上 海 外 国 语 大 学 硕 士 学 位 论 文 对 外 汉 语 初 中 级 副 词 情 境 教 学 研 究 与 实 践 院 系 : 国 际 文 化 交 流 学 院 学 科 专 业 : 汉 语 国 际 教 育 姓 名 : 顾 妍 指 导 教 师 : 缪 俊 2016 年 5 月 Shanghai International Studies University THE STUDY AND PRACTICE

More information

BC04 Module_antenna__ doc

BC04 Module_antenna__ doc http://www.infobluetooth.com TEL:+86-23-68798999 Fax: +86-23-68889515 Page 1 of 10 http://www.infobluetooth.com TEL:+86-23-68798999 Fax: +86-23-68889515 Page 2 of 10 http://www.infobluetooth.com TEL:+86-23-68798999

More information

Microsoft PowerPoint - Aqua-Sim.pptx

Microsoft PowerPoint - Aqua-Sim.pptx Peng Xie, Zhong Zhou, Zheng Peng, Hai Yan, Tiansi Hu, Jun-Hong Cui, Zhijie Shi, Yunsi Fei, Shengli Zhou Underwater Sensor Network Lab 1 Outline Motivations System Overview Aqua-Sim Components Experimental

More information

RPN 2 DeepParts 22 part pool 2 3 HOG 12 LBP 13 Harr - like 15 DPM 16 Deformable Parts Model VGG16 X. Wang 14 VGG Convolutiona

RPN 2 DeepParts 22 part pool 2 3 HOG 12 LBP 13 Harr - like 15 DPM 16 Deformable Parts Model VGG16 X. Wang 14 VGG Convolutiona 2017 22 2 CHINESE JOURNAL OF STEREOLOGY AND IMAGE ANALYSIS Vol. 22 No. 2 June 2017 209 1007-1482 2017 02-0209 -0215 DOI 10. 13505 /j. 1007-1482. 2017. 22. 02. 014 RPN 1 2 3 3 1 1 2 3 1. 100090 2. 100090

More information

% % % % % % ~

% % % % % % ~ 1001-5558 2015 03-0021-16 2010 C91 A 2014 5 2010 N. W. Journal of Ethnology 2015 3 86 2015.No.3 Total No.86 2010 2010 2181.58 882.99 40.47% 1298.59 59.53% 2013 2232.78 847.29 37.95% 1385.49 62.05% 1990

More information

Microsoft PowerPoint SSBSE .ppt [Modo de Compatibilidade]

Microsoft PowerPoint SSBSE .ppt [Modo de Compatibilidade] SSBSE 2015, Bergamo Transformed Search Based Software Engineering: A New Paradigm of SBSE He JIANG, Zhilei Ren, Xiaochen Li, Xiaochen Lai jianghe@dlut.edu.cn School of Software, Dalian Univ. of Tech. Outline

More information

Dan Buettner / /

Dan Buettner / / 39 1 2015 1 Vol. 39 No. 1 January 2015 74 Population Research 80 + /60 + 90 + 90 + 0 80 100028 Measuring and Comparing Population Longevity Level across the Regions of the World Lin Bao Abstract Appropriate

More information

University of Science and Technology of China A dissertation for master s degree Research of e-learning style for public servants under the context of

University of Science and Technology of China A dissertation for master s degree Research of e-learning style for public servants under the context of 中 国 科 学 技 术 大 学 硕 士 学 位 论 文 新 媒 体 环 境 下 公 务 员 在 线 培 训 模 式 研 究 作 者 姓 名 : 学 科 专 业 : 导 师 姓 名 : 完 成 时 间 : 潘 琳 数 字 媒 体 周 荣 庭 教 授 二 一 二 年 五 月 University of Science and Technology of China A dissertation for

More information

IP TCP/IP PC OS µclinux MPEG4 Blackfin DSP MPEG4 IP UDP Winsock I/O DirectShow Filter DirectShow MPEG4 µclinux TCP/IP IP COM, DirectShow I

IP TCP/IP PC OS µclinux MPEG4 Blackfin DSP MPEG4 IP UDP Winsock I/O DirectShow Filter DirectShow MPEG4 µclinux TCP/IP IP COM, DirectShow I 2004 5 IP TCP/IP PC OS µclinux MPEG4 Blackfin DSP MPEG4 IP UDP Winsock I/O DirectShow Filter DirectShow MPEG4 µclinux TCP/IP IP COM, DirectShow I Abstract The techniques of digital video processing, transferring

More information

ENGG1410-F Tutorial 6

ENGG1410-F Tutorial 6 Jianwen Zhao Department of Computer Science and Engineering The Chinese University of Hong Kong 1/16 Problem 1. Matrix Diagonalization Diagonalize the following matrix: A = [ ] 1 2 4 3 2/16 Solution The

More information

* CO3 A 1674-2486 2011 04-0005 - 18 P. 253 * 5 1. 1949 1991 1949 1991 6 2. 7 1 2001 2 2008 8 1 2 2008 11 http / /www. rnd. ncnu. edu. tw /hdcheng /method /ways. doc 2008 / 9 disciplinary matrix 1 1. 2001

More information

A VALIDATION STUDY OF THE ACHIEVEMENT TEST OF TEACHING CHINESE AS THE SECOND LANGUAGE by Chen Wei A Thesis Submitted to the Graduate School and Colleg

A VALIDATION STUDY OF THE ACHIEVEMENT TEST OF TEACHING CHINESE AS THE SECOND LANGUAGE by Chen Wei A Thesis Submitted to the Graduate School and Colleg 上 海 外 国 语 大 学 SHANGHAI INTERNATIONAL STUDIES UNIVERSITY 硕 士 学 位 论 文 MASTER DISSERTATION 学 院 国 际 文 化 交 流 学 院 专 业 汉 语 国 际 教 育 硕 士 题 目 届 别 2010 届 学 生 陈 炜 导 师 张 艳 莉 副 教 授 日 期 2010 年 4 月 A VALIDATION STUDY

More information

穨423.PDF

穨423.PDF Chinese Journal of Science Education 2002,, 423-439 2002, 10(4), 423-439 1 2 1 1 1 2 90 8 10 91 4 9 91 8 22 ) NII 1995 7 14, 1999 1997 (Cooperative Remotely Accessible Learning CORAL) 424 (Collaborative

More information

IPCC CO (IPCC2006) 1 : = ( 1) 1 (kj/kg) (kgc/gj) (tc/t)

IPCC CO (IPCC2006) 1 : = ( 1) 1 (kj/kg) (kgc/gj) (tc/t) 2011 5 5 (278 ) China Industrial Economics May 2011 No.5 1 12 (1. 100005; 2. 066004) [ ] : ; ; : ; ; [ ] ; ; ; [ ]F290 [ ]A [ ]1006-480X(2011)05-0047-11 2008 CO 2 ( ) (2009) (GDP) (Binhocker et al. 2008)

More information

Fig. 1 Frame calculation model 1 mm Table 1 Joints displacement mm

Fig. 1 Frame calculation model 1 mm Table 1 Joints displacement mm 33 2 2011 4 ol. 33 No. 2 Apr. 2011 1002-8412 2011 02-0104-08 1 1 1 2 361003 3. 361009 3 1. 361005 2. GB50023-2009 TU746. 3 A Study on Single-span RC Frame Reinforced with Steel Truss System Yuan Xing-ren

More information

第 2 期 王 向 东 等 : 一 种 运 动 轨 迹 引 导 下 的 举 重 视 频 关 键 姿 态 提 取 方 法 257 竞 技 体 育 比 赛 越 来 越 激 烈, 为 了 提 高 体 育 训 练 的 效 率, 有 必 要 在 体 育 训 练 中 引 入 科 学 定 量 的 方 法 许 多

第 2 期 王 向 东 等 : 一 种 运 动 轨 迹 引 导 下 的 举 重 视 频 关 键 姿 态 提 取 方 法 257 竞 技 体 育 比 赛 越 来 越 激 烈, 为 了 提 高 体 育 训 练 的 效 率, 有 必 要 在 体 育 训 练 中 引 入 科 学 定 量 的 方 法 许 多 2014 年 4 月 图 学 学 报 April 2014 第 35 卷 第 2 期 JOURNAL OF GRAPHICS Vol.35 No.2 一 种 运 动 轨 迹 引 导 下 的 举 重 视 频 关 键 姿 态 提 取 方 法 王 向 东 1, 张 静 文 2, 毋 立 芳 2, 徐 文 泉 (1. 国 家 体 育 总 局 体 育 科 学 研 究 所, 北 京 100061;2. 北 京

More information

1 引言

1 引言 P P 第 40 卷 Vol.40 第 7 期 No.7 计 算 机 工 程 Computer Engineering 014 年 7 月 July 014 开 发 研 究 与 工 程 应 用 文 章 编 号 :1000-348(014)07-081-05 文 献 标 识 码 :A 中 图 分 类 号 :TP391.41 摘 基 于 图 像 识 别 的 震 象 云 地 震 预 测 方 法 谢 庭,

More information

國立中山大學學位論文典藏.PDF

國立中山大學學位論文典藏.PDF The Study on the New Pension Scheme for Civil Servants Evidence from Kaohsiung County I II 1. III Thesis Abstract Title of Thesis The Study on the New Pension Scheme for Civil Servants: Evidence from Kaohsiung

More information

南華大學數位論文

南華大學數位論文 南華大學 碩士論文 中華民國九十五年六月十四日 Elfin Excel I II III ABSTRACT Since Ming Hwa Yuan Taiwanese Opera Company started to cooperate with the Chinese orchestra, the problem of how the participation of Chinese music

More information

:1949, 1936, 1713 %, 63 % (, 1957, 5 ), :?,,,,,, (,1999, 329 ),,,,,,,,,, ( ) ; ( ), 1945,,,,,,,,, 100, 1952,,,,,, ,, :,,, 1928,,,,, (,1984, 109

:1949, 1936, 1713 %, 63 % (, 1957, 5 ), :?,,,,,, (,1999, 329 ),,,,,,,,,, ( ) ; ( ), 1945,,,,,,,,, 100, 1952,,,,,, ,, :,,, 1928,,,,, (,1984, 109 2006 9 1949 3 : 1949 2005, : 1949 1978, ; 1979 1997, ; 1998 2005,,, :,,, 1949, :, ;,,,, 50, 1952 1957 ; ; 60 ; 1978 ; 2003,,,,,,, 1953 1978 1953 1978,,,, 100,,,,, 3,, :100836, :wulijjs @263. net ;,, :

More information

世新稿件end.doc

世新稿件end.doc Research Center For Taiwan Economic Development (RCTED) 2003 8 1 2 Study of Operational Strategies on Biotechnology Pharmaceutical Products Industry in Taiwan -- Case Study on Sinphar Pharmaceutical Company

More information

Microsoft PowerPoint - ATF2015.ppt [相容模式]

Microsoft PowerPoint - ATF2015.ppt [相容模式] Improving the Video Totalized Method of Stopwatch Calibration Samuel C.K. Ko, Aaron Y.K. Yan and Henry C.K. Ma The Government of Hong Kong Special Administrative Region (SCL) 31 Oct 2015 1 Contents Introduction

More information

y 1 = 槡 P 1 1h T 1 1f 1 s 1 + 槡 P 1 2g T 1 2 interference 2f 2 s y 2 = 槡 P 2 2h T 2 2f 2 s 2 + 槡 P 2 1g T 2 1 interference 1f 1 s + n n

y 1 = 槡 P 1 1h T 1 1f 1 s 1 + 槡 P 1 2g T 1 2 interference 2f 2 s y 2 = 槡 P 2 2h T 2 2f 2 s 2 + 槡 P 2 1g T 2 1 interference 1f 1 s + n n 37 1 Vol 37 No 1 2013 1 Journal of Jiangxi Normal UniversityNatural Science Jan 2013 1000-5862201301-0037-05 MISO 郭荣新, 袁继昌 361021 2 RVQ 2 MISO 3 TN 911 7 A 0 MIMO 2 MISO 3 MIMOnetwork MIMO 3GPP LTE-A 2

More information

Microsoft Word - TIP006SCH Uni-edit Writing Tip - Presentperfecttenseandpasttenseinyourintroduction readytopublish

Microsoft Word - TIP006SCH Uni-edit Writing Tip - Presentperfecttenseandpasttenseinyourintroduction readytopublish 我 难 度 : 高 级 对 们 现 不 在 知 仍 道 有 听 影 过 响 多 少 那 次 么 : 研 英 究 过 文 论 去 写 文 时 作 的 表 技 引 示 巧 言 事 : 部 情 引 分 发 言 该 生 使 在 中 用 过 去, 而 现 在 完 成 时 仅 表 示 事 情 发 生 在 过 去, 并 的 哪 现 种 在 时 完 态 成 呢 时? 和 难 过 道 去 不 时 相 关? 是 所 有

More information

~ ~

~ ~ * 40 4 2016 7 Vol. 40 No. 4 July 2016 35 Population Research 2014 1 2016 2016 9101. 0 40 49. 6% 2017 ~ 2021 1719. 5 160 ~ 470 100872 Accumulated Couples and Extra Births under the Universal Tw o-child

More information

Microsoft PowerPoint - Performance Analysis of Video Streaming over LTE using.pptx

Microsoft PowerPoint - Performance Analysis of Video Streaming over LTE using.pptx ENSC 427 Communication Networks Spring 2016 Group #2 Project URL: http://www.sfu.ca/~rkieu/ensc427_project.html Amer, Zargham 301149920 Kieu, Ritchie 301149668 Xiao, Lei 301133381 1 Roadmap Introduction

More information

STEAM STEAM STEAM ( ) STEAM STEAM ( ) 1977 [13] [10] STEM STEM 2. [11] [14] ( )STEAM [15] [16] STEAM [12] ( ) STEAM STEAM [17] STEAM STEAM STEA

STEAM STEAM STEAM ( ) STEAM STEAM ( ) 1977 [13] [10] STEM STEM 2. [11] [14] ( )STEAM [15] [16] STEAM [12] ( ) STEAM STEAM [17] STEAM STEAM STEA 2017 8 ( 292 ) DOI:10.13811/j.cnki.eer.2017.08.017 STEAM 1 1 2 3 4 (1. 130117; 2. + 130117; 3. 130022;4. 518100) [ ] 21 STEAM STEAM STEAM STEAM STEAM STEAM [ ] STEAM ; ; [ ] G434 [ ] A [ ] (1970 ) E-mail:ddzhou@nenu.edu.cn

More information

% % 99% Sautman B. Preferential Policies for Ethnic Minorities in China The Case

% % 99% Sautman B. Preferential Policies for Ethnic Minorities in China The Case 1001-5558 2015 03-0037-11 2000 2010 C95 DOI:10.16486/j.cnki.62-1035/d.2015.03.005 A 1 2014 14CRK014 2013 13SHC012 1 47 2181 N. W. Journal of Ethnology 2015 3 86 2015.No.3 Total No.86 20 70 122000 2007

More information

Vol. 22 No. 4 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Aug GPS,,, : km, 2. 51, , ; ; ; ; DOI: 10.

Vol. 22 No. 4 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Aug GPS,,, : km, 2. 51, , ; ; ; ; DOI: 10. 22 4 2017 8 Vol. 22 No. 4 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Aug. 2017 150080 GPS,,, : 27. 36 km, 2. 51, 110. 43, ; ; ; ; DOI: 10. 15938 /j. jhust. 2017. 04. 015 U469. 13 A 1007-2683

More information

國立中山大學學位論文典藏

國立中山大學學位論文典藏 I II III IV The theories of leadership seldom explain the difference of male leaders and female leaders. Instead of the assumption that the leaders leading traits and leading styles of two sexes are the

More information

<4D6963726F736F667420576F7264202D20D0ECB7C9D4C6A3A8C5C5B0E6A3A92E646F63>

<4D6963726F736F667420576F7264202D20D0ECB7C9D4C6A3A8C5C5B0E6A3A92E646F63> 硕 士 专 业 学 位 论 文 论 文 题 目 性 灵 文 学 思 想 与 高 中 作 文 教 学 研 究 生 姓 名 指 导 教 师 姓 名 专 业 名 称 徐 飞 云 卞 兆 明 教 育 硕 士 研 究 方 向 学 科 教 学 ( 语 文 ) 论 文 提 交 日 期 2012 年 9 月 性 灵 文 学 思 想 与 高 中 作 文 教 学 中 文 摘 要 性 灵 文 学 思 想 与 高 中 作

More information

Preface This guide is intended to standardize the use of the WeChat brand and ensure the brand's integrity and consistency. The guide applies to all d

Preface This guide is intended to standardize the use of the WeChat brand and ensure the brand's integrity and consistency. The guide applies to all d WeChat Search Visual Identity Guidelines WEDESIGN 2018. 04 Preface This guide is intended to standardize the use of the WeChat brand and ensure the brand's integrity and consistency. The guide applies

More information

θ 1 = φ n -n 2 2 n AR n φ i = 0 1 = a t - θ θ m a t-m 3 3 m MA m 1. 2 ρ k = R k /R 0 5 Akaike ρ k 1 AIC = n ln δ 2

θ 1 = φ n -n 2 2 n AR n φ i = 0 1 = a t - θ θ m a t-m 3 3 m MA m 1. 2 ρ k = R k /R 0 5 Akaike ρ k 1 AIC = n ln δ 2 35 2 2012 2 GEOMATICS & SPATIAL INFORMATION TECHNOLOGY Vol. 35 No. 2 Feb. 2012 1 2 3 4 1. 450008 2. 450005 3. 450008 4. 572000 20 J 101 20 ARMA TU196 B 1672-5867 2012 02-0213 - 04 Application of Time Series

More information

Microsoft Word - chnInfoPaper6

Microsoft Word - chnInfoPaper6 文 章 编 号 :3-77(2)-- 文 章 编 号 :92 基 于 中 文 拼 音 输 入 法 数 据 的 汉 语 方 言 词 汇 自 动 识 别 张 燕, 张 扬 2, 孙 茂 松 (. 清 华 大 学 计 算 机 系, 北 京 市 84;2. 搜 狗 科 技 公 司, 北 京 市 84) 摘 要 : 方 言 研 究 领 域 中 的 语 音 研 究 词 汇 研 究 及 语 法 研 究 是 方 言

More information

SVM OA 1 SVM MLP Tab 1 1 Drug feature data quantization table

SVM OA 1 SVM MLP Tab 1 1 Drug feature data quantization table 38 2 2010 4 Journal of Fuzhou University Natural Science Vol 38 No 2 Apr 2010 1000-2243 2010 02-0213 - 06 MLP SVM 1 1 2 1 350108 2 350108 MIP SVM OA MLP - SVM TP391 72 A Research of dialectical classification

More information

a b

a b 38 3 2014 5 Vol. 38 No. 3 May 2014 55 Population Research + + 3 100038 A Study on Implementation of Residence Permit System Based on Three Local Cases of Shanghai Chengdu and Zhengzhou Wang Yang Abstract

More information

Microsoft Word - KSAE06-S0262.doc

Microsoft Word - KSAE06-S0262.doc Stereo Vision based Forward Collision Warning and Avoidance System Yunhee LeeByungjoo KimHogi JungPaljoo Yoon Central R&D Center, MANDO Corporation, 413-5, Gomae-Ri, Gibeung-Eub, Youngin-Si, Kyonggi-Do,

More information

Microsoft Word - 刘 慧 板.doc

Microsoft Word - 刘  慧 板.doc 中 国 环 境 科 学 2012,32(5):933~941 China Environmental Science 系 统 动 力 学 在 空 港 区 域 规 划 环 境 影 响 评 价 中 的 应 用 刘 慧 1,2, 郭 怀 成 1*, 盛 虎 1, 都 小 尚 1,3, 李 娜 1 1, 杨 永 辉 (1. 北 京 大 学 环 境 科 学 与 工 程 学 院, 北 京 100871; 2.

More information

标题

标题 第 19 卷 摇 第 4 期 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 模 式 识 别 与 人 工 智 能 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 Vol. 19 摇 No. 4 摇 006 年 8 月 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 PR & AI 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 摇 Aug 摇 摇

More information

McGraw-Hill School Education Group Physics : Principles and Problems G S 24

McGraw-Hill School Education Group Physics : Principles and Problems G S 24 2017 4 357 GLOBAL EDUCATION Vol. 46 No4, 2017 * 1 / 400715 / 400715 / 400715 1 2010-2020 2 * mjzxzd1401 2012 AHA120008 1 23 3 4-7 8 9 McGraw-Hill School Education Group Physics : Principles and Problems

More information

be invested on the desilting of water sources and to paved canals with cement mortar while drinking water project can focus on the improvement of wate

be invested on the desilting of water sources and to paved canals with cement mortar while drinking water project can focus on the improvement of wate 2011 9 100101 101 3 2007 2000 10 F303 A 1002-9753 2011 09-0030 - 11 Trends on Public Infrastructure and Public Investment Priority in Rural China LUO Ren - fu ZHANG Lin - xiu ZHAO Qi - ran HUANG Ji - kun

More information

Microsoft Word - A200810-897.doc

Microsoft Word - A200810-897.doc 基 于 胜 任 特 征 模 型 的 结 构 化 面 试 信 度 和 效 度 验 证 张 玮 北 京 邮 电 大 学 经 济 管 理 学 院, 北 京 (100876) E-mail: weeo1984@sina.com 摘 要 : 提 高 结 构 化 面 试 信 度 和 效 度 是 面 试 技 术 研 究 的 核 心 内 容 近 年 来 国 内 有 少 数 学 者 探 讨 过 基 于 胜 任 特 征

More information

WTO

WTO 10384 200015128 UDC Exploration on Design of CIB s Human Resources System in the New Stage (MBA) 2004 2004 2 3 2004 3 2 0 0 4 2 WTO Abstract Abstract With the rapid development of the high and new technique

More information

Outline Speech Signals Processing Dual-Tone Multifrequency Signal Detection 云南大学滇池学院课程 : 数字信号处理 Applications of Digital Signal Processing 2

Outline Speech Signals Processing Dual-Tone Multifrequency Signal Detection 云南大学滇池学院课程 : 数字信号处理 Applications of Digital Signal Processing 2 CHAPTER 10 Applications of Digital Signal Processing Wang Weilian wlwang@ynu.edu.cn School of Information Science and Technology Yunnan University Outline Speech Signals Processing Dual-Tone Multifrequency

More information

I

I The Effect of Guided Discovery on The Learning Achievement and Learning Transfer of Grade 5 Students in Primary Schools I II Abstract The Effect of Guided Discovery on The Learning Achievement And Learning

More information

Chinese Journal of Applied Probability and Statistics Vol.25 No.4 Aug (,, ;,, ) (,, ) 应用概率统计 版权所有, Zhang (2002). λ q(t)

Chinese Journal of Applied Probability and Statistics Vol.25 No.4 Aug (,, ;,, ) (,, ) 应用概率统计 版权所有, Zhang (2002). λ q(t) 2009 8 Chinese Journal of Applied Probability and Statistics Vol.25 No.4 Aug. 2009,, 541004;,, 100124),, 100190), Zhang 2002). λ qt), Kolmogorov-Smirov, Berk and Jones 1979). λ qt).,,, λ qt),. λ qt) 1,.

More information

Microsoft PowerPoint - talk8.ppt

Microsoft PowerPoint - talk8.ppt Adaptive Playout Scheduling Using Time-scale Modification Yi Liang, Nikolaus Färber Bernd Girod, Balaji Prabhakar Outline QoS concerns and tradeoffs Jitter adaptation as a playout scheduling scheme Packet

More information

致 谢 本 人 自 2008 年 6 月 从 上 海 外 国 语 大 学 毕 业 之 后, 于 2010 年 3 月 再 次 进 入 上 外, 非 常 有 幸 成 为 汉 语 国 际 教 育 专 业 的 研 究 生 回 顾 三 年 以 来 的 学 习 和 生 活, 顿 时 感 觉 这 段 时 间 也

致 谢 本 人 自 2008 年 6 月 从 上 海 外 国 语 大 学 毕 业 之 后, 于 2010 年 3 月 再 次 进 入 上 外, 非 常 有 幸 成 为 汉 语 国 际 教 育 专 业 的 研 究 生 回 顾 三 年 以 来 的 学 习 和 生 活, 顿 时 感 觉 这 段 时 间 也 精 英 汉 语 和 新 实 用 汉 语 课 本 的 对 比 研 究 The Comparative Study of Jing Ying Chinese and The New Practical Chinese Textbook 专 业 : 届 别 : 姓 名 : 导 师 : 汉 语 国 际 教 育 2013 届 王 泉 玲 杨 金 华 1 致 谢 本 人 自 2008 年 6 月 从 上 海 外

More information

#4 ~ #5 12 m m m 1. 5 m # m mm m Z4 Z5

#4 ~ #5 12 m m m 1. 5 m # m mm m Z4 Z5 2011 6 6 153 JOURNAL OF RAILWAY ENGINEERING SOCIETY Jun 2011 NO. 6 Ser. 153 1006-2106 2011 06-0014 - 07 300142 ABAQUS 4. 287 mm 6. 651 mm U455. 43 A Analysis of Impact of Shield Tunneling on Displacement

More information

untitled

untitled Co-integration and VECM Yi-Nung Yang CYCU, Taiwan May, 2012 不 列 1 Learning objectives Integrated variables Co-integration Vector Error correction model (VECM) Engle-Granger 2-step co-integration test Johansen

More information

度 身 體 活 動 量 ; 芬 蘭 幼 兒 呈 現 中 度 身 體 活 動 量 之 比 例 高 於 臺 灣 幼 兒 (5) 幼 兒 在 投 入 度 方 面 亦 達 顯 著 差 異 (χ²=185.35, p <.001), 芬 蘭 與 臺 灣 幼 兒 多 半 表 現 出 中 度 投 入 與 高 度

度 身 體 活 動 量 ; 芬 蘭 幼 兒 呈 現 中 度 身 體 活 動 量 之 比 例 高 於 臺 灣 幼 兒 (5) 幼 兒 在 投 入 度 方 面 亦 達 顯 著 差 異 (χ²=185.35, p <.001), 芬 蘭 與 臺 灣 幼 兒 多 半 表 現 出 中 度 投 入 與 高 度 臺 灣 與 芬 蘭 幼 兒 園 室 內 自 由 遊 戲 內 涵 之 探 討 林 昭 溶 毛 萬 儀 經 國 管 理 暨 健 康 學 院 幼 兒 保 育 系 副 教 授 joyce@ems.cku.edu.tw 吳 敏 而 國 家 教 育 研 究 院 研 究 員 rozwu@mail.naer.edu.tw wanyi@ems.cku.edu.tw 摘 要 自 由 遊 戲 被 視 為 是 幼 兒 的

More information

Microsoft Word - ED-774.docx

Microsoft Word - ED-774.docx journal.newcenturyscience.com/index.php/gjanp Global Journal of Advanced Nursing Practice,214,Vol.1,No.1 The practicality of an improved method of intravenous infusion exhaust specialized in operating

More information

http / /yxxy. cbpt. cnki. net / % % %

http / /yxxy. cbpt. cnki. net / % % % 2017 3 Mar. 2017 5 2 Chongqing Higher Education Research Vol. 5 No. 2 DOI 10. 15998 /j. cnki. issn1673-8012. 2017. 02. 006 230039 2011 2015 2016 G649. 21 A 1673-8012 2017 02-0037-11 2017-01-03 2015zdjy024

More information

(Pattern Recognition) 1 1. CCD

(Pattern Recognition) 1 1. CCD ********************************* ********************************* (Pattern Recognition) 1 1. CCD 2. 3. 4. 1 ABSTRACT KeywordsMachine Vision, Real Time Inspection, Image Processing The purpose of this

More information

10384 200115009 UDC Management Buy-outs MBO MBO MBO 2002 MBO MBO MBO MBO 000527 MBO MBO MBO MBO MBO MBO MBO MBO MBO MBO MBO Q MBO MBO MBO Abstract Its related empirical study demonstrates a remarkable

More information

2005 5,,,,,,,,,,,,,,,,, , , 2174, 7014 %, % 4, 1961, ,30, 30,, 4,1976,627,,,,, 3 (1993,12 ),, 2

2005 5,,,,,,,,,,,,,,,,, , , 2174, 7014 %, % 4, 1961, ,30, 30,, 4,1976,627,,,,, 3 (1993,12 ),, 2 3,,,,,, 1872,,,, 3 2004 ( 04BZS030),, 1 2005 5,,,,,,,,,,,,,,,,, 1928 716,1935 6 2682 1928 2 1935 6 1966, 2174, 7014 %, 94137 % 4, 1961, 59 1929,30, 30,, 4,1976,627,,,,, 3 (1993,12 ),, 2 , :,,,, :,,,,,,

More information

% % 34

% % 34 * 2000 2005 1% 1% 1% 1% * VZDA2010-15 33 2011. 3 2009 2009 2004 2008 1982 1990 2000 2005 1% 1 1 2005 1% 34 2000 2005 1% 35 2011. 3 2000 0. 95 20-30 209592 70982 33. 9% 2005 1% 258 20-30 372301 115483 31.

More information

(baking powder) 1 ( ) ( ) 1 10g g (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal D

(baking powder) 1 ( ) ( ) 1 10g g (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal D ( ) 4 1 1 1 145 1 110 1 (baking powder) 1 ( ) ( ) 1 10g 1 1 2.5g 1 1 1 1 60 10 (two level design, D-optimal) 32 1/2 fraction Two Level Fractional Factorial Design D-Optimal Design 1. 60 120 2. 3. 40 10

More information

~ 10 2 P Y i t = my i t W Y i t 1000 PY i t Y t i W Y i t t i m Y i t t i 15 ~ 49 1 Y Y Y 15 ~ j j t j t = j P i t i = 15 P n i t n Y

~ 10 2 P Y i t = my i t W Y i t 1000 PY i t Y t i W Y i t t i m Y i t t i 15 ~ 49 1 Y Y Y 15 ~ j j t j t = j P i t i = 15 P n i t n Y * 35 4 2011 7 Vol. 35 No. 4 July 2011 3 Population Research 1950 ~ 1981 The Estimation Method and Its Application of Cohort Age - specific Fertility Rates Wang Gongzhou Hu Yaoling Abstract Based on the

More information

59-81

59-81 BIBLID 0254-4466(2001)19:2 pp. 59-81 19 2 90 12 * 59 60 19 2 1498-1583 6 1572 12 27 1525-1582 1572-1620 1368-1398 1426-1435 1450-1456 1610-1695 15 1538-1588 1535-1608 61 1 1503-1583 1516-1591 1472-1528

More information

相 關 技 術, 在 裝 置 上 創 造 出 一 個 令 人 驚 豔 虛 擬 的 幻 境 ; 除 此 之 外, 還 能 與 虛 擬 出 來 的 物 件 進 行 互 動, 已 陸 陸 續 續 被 應 用 在 教 育 研 究 娛 樂 生 活 等 各 個 方 面 認 知 風 格 (Cognitive St

相 關 技 術, 在 裝 置 上 創 造 出 一 個 令 人 驚 豔 虛 擬 的 幻 境 ; 除 此 之 外, 還 能 與 虛 擬 出 來 的 物 件 進 行 互 動, 已 陸 陸 續 續 被 應 用 在 教 育 研 究 娛 樂 生 活 等 各 個 方 面 認 知 風 格 (Cognitive St 擴 增 實 境 互 動 遊 戲 於 認 知 風 格 之 評 估 The Study on Cognitive Styles and a Mobile Augmented Reality Interactive Game 謝 旻 儕 1*, 蔡 孟 君 1 2, 林 豪 鏘 1 和 春 技 術 學 院 資 訊 管 理 系 2 國 立 臺 南 大 學 數 位 學 習 科 技 學 系 * shiehminchai@gmail.com

More information

LH_Series_Rev2014.pdf

LH_Series_Rev2014.pdf REMINDERS Product information in this catalog is as of October 2013. All of the contents specified herein are subject to change without notice due to technical improvements, etc. Therefore, please check

More information

本 論 文 獲 行 政 院 客 家 委 員 會 99 年 客 家 研 究 優 良 博 碩 士 論 文 獎 助

本 論 文 獲 行 政 院 客 家 委 員 會 99 年 客 家 研 究 優 良 博 碩 士 論 文 獎 助 國 立 雲 林 科 技 大 學 設 計 運 算 研 究 所 碩 士 班 碩 士 論 文 人 物 類 數 位 典 藏 加 值 應 用 之 研 究 - 以 桃 園 縣 客 家 文 化 館 之 鄧 雨 賢 文 化 加 值 產 品 設 計 為 例 A Study of Value-added Applications in Digital Archiving: Deng Yu-Shian's memorabilia

More information

考試學刊第10期-內文.indd

考試學刊第10期-內文.indd misconception 101 Misconceptions and Test-Questions of Earth Science in Senior High School Chun-Ping Weng College Entrance Examination Center Abstract Earth Science is a subject highly related to everyday

More information

01何寄澎.doc

01何寄澎.doc 1 * ** * ** 2003 11 1-36 Imitation and the Formation and Interpretation of the Canon: Lu Chi s Nigushi Ho Chi-p eng Professor, Department of Chinese Literature, National Taiwan University. Hsu Ming-ch

More information

論 文 摘 要 本 文 乃 係 兩 岸 稅 務 爭 訟 制 度 之 研 究, 蓋 稅 務 爭 訟 在 行 訴 訟 中 一 直 占 有 相 當 高 的 比 例, 惟 其 勝 訴 率 一 直 偏 低, 民 87 年 10 月 28 日 行 訴 訟 法 經 幅 修 正 後, 審 級 部 分 由 一 級 一

論 文 摘 要 本 文 乃 係 兩 岸 稅 務 爭 訟 制 度 之 研 究, 蓋 稅 務 爭 訟 在 行 訴 訟 中 一 直 占 有 相 當 高 的 比 例, 惟 其 勝 訴 率 一 直 偏 低, 民 87 年 10 月 28 日 行 訴 訟 法 經 幅 修 正 後, 審 級 部 分 由 一 級 一 法 院 碩 士 在 職 專 班 碩 士 論 文 指 導 教 授 : 王 文 杰 博 士 兩 岸 稅 務 爭 訟 制 度 之 比 較 研 究 A comparative study on the system of cross-straits tax litigation 研 究 生 : 羅 希 寧 中 華 民 一 0 一 年 七 月 論 文 摘 要 本 文 乃 係 兩 岸 稅 務 爭 訟 制 度 之

More information

Microsoft PowerPoint - ryz_030708_pwo.ppt

Microsoft PowerPoint - ryz_030708_pwo.ppt Long Term Recovery of Seven PWO Crystals Ren-yuan Zhu California Institute of Technology CMS ECAL Week, CERN Introduction 20 endcap and 5 barrel PWO crystals went through (1) thermal annealing at 200 o

More information

JOURNAL OF EARTHQUAKE ENGINEERING AND ENGINEERING VIBRATION Vol. 31 No. 5 Oct /35 TU3521 P315.

JOURNAL OF EARTHQUAKE ENGINEERING AND ENGINEERING VIBRATION Vol. 31 No. 5 Oct /35 TU3521 P315. 31 5 2011 10 JOURNAL OF EARTHQUAKE ENGINEERING AND ENGINEERING VIBRATION Vol. 31 No. 5 Oct. 2011 1000-1301 2011 05-0075 - 09 510405 1 /35 TU3521 P315. 8 A Earthquake simulation shaking table test and analysis

More information

K301Q-D VRT中英文说明书141009

K301Q-D VRT中英文说明书141009 THE INSTALLING INSTRUCTION FOR CONCEALED TANK Important instuction:.. Please confirm the structure and shape before installing the toilet bowl. Meanwhile measure the exact size H between outfall and infall

More information

P

P 100871 I0-03 A 1671-7511 2011 04-0072 - 11 6 18 1 2 3 4 2010-01 - 04 1 2 3 4 72 1 20 2 1 P309 310 1 2 73 1 3 2 1933 1 2 3 74 13 1976 7 2 1 2 P281 1929 1930 1935 1937 1948 1 2 1933 3 25 11 1998 5 ~ 1999

More information

1556 地 理 科 学 进 展 30 卷 他 关 于 农 村 住 房 结 构 与 抗 震 性 能 的 研 究, 则 多 是 从 工 程 抗 灾 的 角 度, 研 究 某 种 构 造 类 型 的 房 屋, 力 图 找 到 传 统 房 屋 的 结 构 失 误 和 新 建 房 屋 中 存 在 的 问 [

1556 地 理 科 学 进 展 30 卷 他 关 于 农 村 住 房 结 构 与 抗 震 性 能 的 研 究, 则 多 是 从 工 程 抗 灾 的 角 度, 研 究 某 种 构 造 类 型 的 房 屋, 力 图 找 到 传 统 房 屋 的 结 构 失 误 和 新 建 房 屋 中 存 在 的 问 [ 第 30 卷 第 12 期 2011 年 12 月 地 理 科 学 进 展 PROGRESS IN GEOGRAPHY Vol.30, No.12 Dec., 2012 中 国 农 村 地 区 住 房 结 构 的 区 域 差 异 及 其 影 响 因 素 1, 2, 李 方 一 3 1,, 高 晓 路 2 2, 王 英 杰 (1. 中 国 科 学 院 区 域 可 持 续 发 展 分 析 与 模 拟 重

More information

D A

D A 2015 4 D822.333 A 0452 8832 2015 4 0014-12 14 The Second ASEAN Regional Forum: The ASEAN Regional Forum, A Concept Paper, in ASEAN Regional Forum Documents Series 1994-2006, ASEAN Secretariat, Jakarta,

More information

國立中山大學學位論文典藏.PDF

國立中山大學學位論文典藏.PDF 中 國 文 學 系 國 立 中 山 大 學, 碩 士 論 文 國 立 中 山 大 學 中 國 文 學 系 碩 士 論 文 Department of Chinese Literature 肉 蒲 團 研 究 National Sun Yat-sen University Master Thesis 肉 蒲 團 研 究 The Research of Rou Pu Tuan 研 究 生 : 林 欣 穎

More information

前 言 香 港 中 文 大 學 優 質 學 校 改 進 計 劃 ( 下 稱 計 劃 ) 團 隊 自 1998 年 起 積 極 於 本 地 推 動 理 論 及 實 踐 並 重 的 學 校 改 進 工 作, 並 逐 步 發 展 成 為 本 地 最 具 規 模 的 校 本 支 援 服 務 品 牌, 曾 支

前 言 香 港 中 文 大 學 優 質 學 校 改 進 計 劃 ( 下 稱 計 劃 ) 團 隊 自 1998 年 起 積 極 於 本 地 推 動 理 論 及 實 踐 並 重 的 學 校 改 進 工 作, 並 逐 步 發 展 成 為 本 地 最 具 規 模 的 校 本 支 援 服 務 品 牌, 曾 支 香 港 中 文 大 學 香 港 教 育 研 究 所 優 質 學 校 改 進 計 劃 : 學 習 差 異 支 援 及 學 校 起 動 計 劃 聯 合 主 辦 中 學 聯 校 教 師 專 業 發 展 日 主 題 : 照 顧 學 習 差 異 日 期 :2011 年 12 月 9 日 ( 星 期 五 ) 時 間 : 上 午 9 時 正 至 下 午 4 時 15 分 地 點 : 樂 善 堂 余 近 卿 中 學

More information

天 主 教 輔 仁 大 學 社 會 學 系 學 士 論 文 小 別 勝 新 婚? 久 別 要 離 婚? 影 響 遠 距 家 庭 婚 姻 感 情 因 素 之 探 討 Separate marital relations are getting better or getting worse? -Exp

天 主 教 輔 仁 大 學 社 會 學 系 學 士 論 文 小 別 勝 新 婚? 久 別 要 離 婚? 影 響 遠 距 家 庭 婚 姻 感 情 因 素 之 探 討 Separate marital relations are getting better or getting worse? -Exp 天 主 教 輔 仁 大 學 社 會 學 系 學 士 論 文 小 別 勝 新 婚? 久 別 要 離 婚? 影 響 遠 距 家 庭 婚 姻 感 情 因 素 之 探 討 Separate marital relations are getting better or getting worse? -Explore the impact of emotional factors couples do not

More information

28 2015 3 1 2 2011-2012 38 985 31 EXCEL 2005-2010 985 3 985 2011 2012 EXCEL 38 985 31 985 2011-2012 38 985 1. 38 985 90 20 70 1 2005-2010 2011-2012 13

28 2015 3 1 2 2011-2012 38 985 31 EXCEL 2005-2010 985 3 985 2011 2012 EXCEL 38 985 31 985 2011-2012 38 985 1. 38 985 90 20 70 1 2005-2010 2011-2012 13 36 3 Vol. 36 No.3 2 0 1 5 5 May 2 0 1 5 985 2011-2012 361005 2011-2012 985 985 985 2005-2012 985 985 G645 A 1001-4519 2015 03-0027 - 12 DOI 10. 14138 /j. 1001-4519. 2015. 03. 002712 2005-2010 985 2012

More information

RESEARCH ON HIGHER EDUCATION Number 4, 2013(General Serial No.78) CONTENTS Colleges and Universities Forum Three-Year Blueprint of Undergraduate Cours

RESEARCH ON HIGHER EDUCATION Number 4, 2013(General Serial No.78) CONTENTS Colleges and Universities Forum Three-Year Blueprint of Undergraduate Cours RESEARCH ON HIGHER EDUCATION 目 录 高 校 论 坛 李 晓 慧 等 北 京 服 装 学 院 本 科 教 学 与 人 才 培 养 三 年 设 想 3 国 家 级 实 验 教 学 示 范 中 心 建 设 赵 平 以 创 新 型 人 才 培 养 为 导 向 构 建 服 装 服 饰 实 践 教 学 平 台 4 王 永 进 服 装 专 业 实 验 室 建 设 与 实 验 教 学

More information

% GIS / / Fig. 1 Characteristics of flood disaster variation in suburbs of Shang

% GIS / / Fig. 1 Characteristics of flood disaster variation in suburbs of Shang 20 6 2011 12 JOURNAL OF NATURAL DISASTERS Vol. 20 No. 6 Dec. 2011 1004-4574 2011 06-0094 - 05 200062 1949-1990 1949 1977 0. 8 0. 03345 0. 01243 30 100 P426. 616 A Risk analysis of flood disaster in Shanghai

More information

untitled

untitled Improving Students' English Learning In The Institute of Technical & Vocational Education By The Strategies of TQM A Case Study of South Taiwan University of Technology NSC 93-2516-S-218-003 930801 950731

More information

62 戲劇學刊 An Analysis of Cao-xie-gong Zhen in Tainan Abstract Te-yu Shih* Tainan is among the first areas that were developed in Taiwan, and there are a

62 戲劇學刊 An Analysis of Cao-xie-gong Zhen in Tainan Abstract Te-yu Shih* Tainan is among the first areas that were developed in Taiwan, and there are a 61 * 1870 61-85 TAIPEI THEATRE JOURNAL 27 (2018): 61-85 School of Theatre Arts, Taipei National University of the Arts 2017.12.21 2018.1.11 * 62 戲劇學刊 An Analysis of Cao-xie-gong Zhen in Tainan Abstract

More information

: (2012) Control Theory & Applications Vol. 29 No. 1 Jan Dezert-Smarandache 1,2, 2,3, 2 (1., ; 2., ;

: (2012) Control Theory & Applications Vol. 29 No. 1 Jan Dezert-Smarandache 1,2, 2,3, 2 (1., ; 2., ; 29 1 2012 1 : 1000 8152(2012)01 0079 06 Control Theory & Applications Vol. 29 No. 1 Jan. 2012 Dezert-Smarandache 1,2, 2,3, 2 (1., 102249; 2., 264001; 3., 410073) :, Dezert-Smarandache (DSmT),. DSmT 3 :.,,

More information

赵 鹍 : 持 久 的 卓 越 : 日 本 年 PISA 测 试 结 果 的 影 响 因 素 分 析 及 启 示 本 的 正 常 社 会 生 活 也 深 受 教 育 影 响 日 本 的 报 刊 文 章 通 常 认 为, 读 者 能 看 懂 复 杂 的 统 计 表, 能 理 解 有

赵 鹍 : 持 久 的 卓 越 : 日 本 年 PISA 测 试 结 果 的 影 响 因 素 分 析 及 启 示 本 的 正 常 社 会 生 活 也 深 受 教 育 影 响 日 本 的 报 刊 文 章 通 常 认 为, 读 者 能 看 懂 复 杂 的 统 计 表, 能 理 解 有 外 国 中 小 学 教 育 2012 年 第 11 期 持 久 的 卓 越 : 日 本 2000-2009 年 PISA 测 试 结 果 的 影 响 因 素 分 析 及 启 示 赵 鹍 摘 要 : 日 本 在 2000-2009 年 的 四 轮 国 际 学 生 评 估 项 目 PISA 测 试 中 表 现 出 色, 一 直 位 于 或 接 近 世 界 领 先 水 平 这 一 成 果 与 日 本 的

More information

E I

E I Research on Using Art-play to Construct Elementary School Students' Visual Art Aesthetic Sensibility ~Case of Da-Yuan Elementary School E I E II Abstract Research on Using Art-play to Construct Elementary

More information

Vol. 22 No. 2 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Apr ,,,,, Apriori,,,,,,,,

Vol. 22 No. 2 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Apr ,,,,, Apriori,,,,,,,, 22 2 2017 4 Vol. 22 No. 2 JOURNAL OF HARBIN UNIVERSITY OF SCIENCE AND TECHNOLOGY Apr. 2017 1 2 3 2 1. 150080 2. 150080 3. 150080,,,,, Apriori,,,,,,,, ; ; ; ; DOI: 10. 15938 /j. jhust. 2017. 02. 023 TN911.

More information

9 21-40 2004 12 * * 22 9 1 2 3 1 1992 2 1960 2 3 1984 8 87 23 4 5 1697 AD 1779 6 7 8 9 10 11 12 4 1977 109-112 5 87 41993 13-38 6 614 7 8 632 9 1974 8 10 631 11 12 632 9 24 13 14 13 1990 14 25 15 16 15

More information

Journal of Northwestern Polytechnical University Apr. Vol No. 2 ICM ICM ICM ICM ICM ICM TP A

Journal of Northwestern Polytechnical University Apr. Vol No. 2 ICM ICM ICM ICM ICM ICM TP A 2012 4 30 2 Journal of Northwestern Polytechnical University Apr. Vol. 30 2012 No. 2 ICM 710072 ICM ICM ICM ICM ICM TP391. 41 A 1000-2758 2012 02-0201-05 6 2001 Kinser Johnson ICM intersecting cortical

More information