Microsoft PowerPoint - pvt-lvelho.ppt

Size: px

Start display at page:

Download "Microsoft PowerPoint - pvt-lvelho.ppt"

鸭侯
4 years ago
Views:

The Place Research @ Microsoft Beijing, 2002 / 2003 Luiz Velho IMPA Part 1 2 The People The Work Vision Graphics NPR Multimedia Shum, Harry Guo, Baining Xu, Ying-Qing Zhang, Ya-Qin Yunbo Cao Eric

Gao Xiaoming Gao Song Ge Xiaodong Gu Chuanxiong Guo Zihua Guo Hong Guo Baining Guo Lei He Chun-Hui Hu Yuxiao Hu Yaohua Hu Xian-Sheng Hua Chao Huang Ling Huang Jinxia Huang Chang-Ning Huang Shipeng

Ligang Liu Ce Liu Nien-Chen Liu Yunxin Liu Xinguo Liu Boying Lu Lie Lu Yufei Ma Wei-Ying Ma Xiaoxu Ma Chengyuan Ma Xin Ma Hu Peng Frank Seide Xiao-Li-Sheila Shang Jacky Shen Yu Shi Heung-Yeung Shum

Xing Xie Yongqiang Xiong Mingqing Xu Yihua Xu Ying-Qing Xu Hongyun Yang Yunxia Yang Binbin Yanhong Richard Yuqi Yao Qiufeng Yin Roger Yu Keman Yu Chun Yuan Huajun Zeng Yue Zhang Gao Zhang Bo Zhang

Zou Yu Zou PS: Not just SIGGRAPH papers 3 4 Genealogy of a SIGGRAPH Paper Choose a Hot / New Topic Know the Area Go Beyond the State-of-the-Art Show Convincing (and Impressive) Results Make an

1 The Place Microsoft Beijing, 2002 / 2003 Luiz Velho IMPA Part 1 2 The People The Work Vision Graphics NPR Multimedia Shum, Harry Guo, Baining Xu, Ying-Qing Zhang, Ya-Qin Yunbo Cao Eric Chang George H-G Chen Lei-Eileen Chen Tong-Xian Chen Gang Chen Zheng Chen Liyong Chen Yu Chen Qi Chen Haitao Chen Yanyun Chen Kaijiang Chen Min Chu Stephen Dahl Yingnong Dang Jianfeng Gao Guangping Gao Xiaoming Gao Song Ge Xiaodong Gu Chuanxiong Guo Zihua Guo Hong Guo Baining Guo Lei He Chun-Hui Hu Yuxiao Hu Yaohua Hu Xian-Sheng Hua Chao Huang Ling Huang Jinxia Huang Chang-Ning Huang Shipeng Li Mingjing Li Jiang Li Stan Li Hang Li Barn-Wan Li Jianghong(Angela) Li Mu Li Cong Li Xun Li Yantao Li Lin Liang Hongbin Liao Shiding Lin Steve Lin Zhouchen Lin Bin Lin Zuoyang Liu Xiaobo Liu Ligang Liu Ce Liu Nien-Chen Liu Yunxin Liu Xinguo Liu Boying Lu Lie Lu Yufei Ma Wei-Ying Ma Xiaoxu Ma Chengyuan Ma Xin Ma Hu Peng Frank Seide Xiao-Li-Sheila Shang Jacky Shen Yu Shi Heung-Yeung Shum Hong-Hui Sun Yan-Feng Sun Kun Tan Xin Tong Jian Wang Jin Wang Qing Wang Lifeng Wang Yijin Wang Lei Wang Xianfang Wang Qiang Wang Mingyu Wang Ji-Rong Wen Hua Wu Feng Wu Jiang Wu Zhe Xiang Rong Xiao Xing Xie Yongqiang Xiong Mingqing Xu Yihua Xu Ying-Qing Xu Hongyun Yang Yunxia Yang Binbin Yanhong Richard Yuqi Yao Qiufeng Yin Roger Yu Keman Yu Chun Yuan Huajun Zeng Yue Zhang Gao Zhang Bo Zhang Qian Zhang Ya-Qin Zhang Hong-Jiang Zhang Lei Zhang Yu Zhang Chunhui Zhang Zheng Zhang Ye Zhang Zhunping Zhang Benyu Zhang Yong Zhao Wei Zheng Ming Zhou Jian-Lai Zhou Kun Zhou Bin Zhu Wenwu Zhu Xinli Zou Yu Zou PS: Not just SIGGRAPH papers 3 4 Genealogy of a SIGGRAPH Paper Choose a Hot / New Topic Know the Area Go Beyond the State-of-the-Art Show Convincing (and Impressive) Results Make an Perfect Presentation The Paper Synthesis of Progressively-Variant Textures on Arbitrary Surfaces Jingdan Zhang. Tsinghua Univ. Kun Zhou MSR Asia Luiz Velho IMPA Baining Guo MSR Asia Harry Shum MSR Asia Some Luck Helps! 5 6

The State-of-the-Art Homogeneous Texture Beyond the State of the Art Natural Textures on Complex Objects Synthesis on Surfaces 7 8 Texture Models Julesz Conjecture (global statistics)

System - attentive system: focal attention scanned serially.

) = µ (constant) Var(X s ) = σ 2 (constant) Cov(X s, X r ) = C(r-s) (function of the distance) 9 10 Probabilistic Model Analysis Algorithm Markov Random Field Lattice S = { s = (i; j)

2 The State-of-the-Art Homogeneous Texture Beyond the State of the Art Natural Textures on Complex Objects Synthesis on Surfaces 7 8 Texture Models Julesz Conjecture (global statistics) Textures cannot be spontaneously discriminated if they have the same first-order and second-order statistics Textons (local features) - local conspicuopus features Two-Level Visual System - attentive system: focal attention scanned serially. - pre-attentive system: distributed attention mediated by parallel process Homogeneous Textures Texture Model Local Stationary Second-Order Stationarity Random Process: { X s } E(X s ) = µ (constant) Var(X s ) = σ 2 (constant) Cov(X s, X r ) = C(r-s) (function of the distance) 9 10 Probabilistic Model Analysis Algorithm Markov Random Field Lattice S = { s = (i; j) : 0 i; j < M} Neighborhood N = { N s S, s S} N s s X s Non-Parametric MRF Model Estimation Given a sample image y Ψ and a neighborhood system N, build a multidimensional histogram F LPCDF estimation Local Conditional Probability Density Function P(X s = x s ) = P(x s x r, r N s } ; x s Λ s 1D Example 11 12

3 Synthesis Algorithm Image X Surface Synthesis Sampling from the LPCDF Impractical (Too Expensive, Huge Memory) image mesh Neighborhood Search Causal Neighborhoods Copy Search noise Pixels/Vertices Local Orientation (Vector field) Synthesis Order Neighborhood grid v u scanline mesh re-tiling normal tangent bitangent distance flattening/resampling Quest for More General Textures Progressively-Variant Textures characteristics change smoothly over domain Locally-Stationary Random Fields - Covariance is a smooth function over the domain Cov(X s, X r ) = F(s, r-s ), F of class C k Recent Research (Mallat, Donoho, Dalhaus ) Local Cosine Best Basis Spectral Estimation Need different approach for synthesis 15 Contributions Method for Generating PVT on Surfaces New Texture Model Texton Based New CreationTools (2D) User Control New Algorithm for Synthesis on Surfaces (3D) Robustness 16 The Model How to Create the Model? T P-V = (S, M, F, V) Estimation Techniques (Malik) Analysis from Samples Difficult Problem / Little Control Image Processing Techniques Image Sample Texton Map Transition Function Orientation Field Transformation of Samples 2D Manipulation Tools Interactive Good User Control (our proposal ) 17 18

4 Building the Model Tools Field-Guided Distortion Mask Extraction Feature-Based Warping Adapted Similarity Editing Texton-Based Blending Distortion Input: Homogeneous Sample Transition Function * Orientation Field * Processing 2D Synthesis Output Image Sample Color Thresholding Segmentation Warping Feature Based (Texton Filtering) Hand Painted Weighted Similarity ( extend Brooks ) Transition Function Operations Painting Warping Morphological Editing 1. Generate Transition Texton Map 2. Synthesize Image Sample Blending 23 24

Example Texton Map Blending Crafting a Leopard Skin 1. Interpolate Maps in 3D 2.

Threshold 25 26 Synthesis Overview Input Problem: 2D Texture Model Matching

Texture Breaking Transition / Orientation Fields on Surface User-Specified Our

Mesh Neighborhoods Single Resolution / One Pass Algorithm 1 2 Nc(s, p), Nm(s, p),

5 Example Texton Map Blending Crafting a Leopard Skin 1. Interpolate Maps in 3D 2. Take Diagonal Slice 3. Gaussian Blur 4. Threshold Synthesis Overview Input Problem: 2D Texture Model Matching Transition / Orientation TP-V = (S, M, F, V) Triangulated Surface Challenge Mesh Texture Breaking Transition / Orientation Fields on Surface User-Specified Our Solution: Output New Metric Texture Maps Texture Coordinate Assignement Mesh Neighborhoods Single Resolution / One Pass Algorithm 1 2 Nc(s, p), Nm(s, p), Orientation Alignment, o = F(p) Scale Adjustment, s = V(p) Resampling 3 Resampled Grid 3D Patch Compare 4 Resample Flatten 2D Patch

Progressively-Variant Synthesis Acceleration

Functions Needs Special Acceleration 31 32 Metric

histogram) Local Orientation Control Orientation on

6 Progressively-Variant Synthesis Acceleration Candidate Pool (i.e. sample neighborhoods) K-Coherence Match Transition Functions Needs Special Acceleration Metric Perceptual-Based (Texton) damage resistant (small histogram) Local Orientation Control Orientation on Mesh and Texture w/o texture orientation Obs: 2-pass match Summary Results Warping Giraffe Tiger Horse Blending Venus Bunny Video 35 36

7 Giraffe Tiger Horse Venus Bunny Timings 41 42

8 More Results Effect of Different Texton Maps Comparisons Wei Levoy Algorithm 2D Synthesis Texton Map helps to deal with difficult textures 47 48

9 Blending Comparison Synthesis Failure 53 54

10 Blending Failure Future Work More Control of Feature-Based Blending Multi-Way Transition Other Ways to Control Texture Variation Synthesis of BTFs Micro-Geometry Synthesis To be continued

Mixtions Pin Yin Homepage

Mixtions Pin Yin Homepage an tai yin 安胎饮 775 ba wei dai xia fang 八味带下方 756 ba zhen tang 八珍汤 600 ba zheng san 八正散 601 bai he gu jin tang 百合固金汤 680 bai hu jia ren shen tang 白虎加人参汤 755 bai hu tang 白虎汤 660 bai