Digital Geometry Processing

Size: px

Start display at page:

Download "Digital Geometry Processing"

表七纳喇
5 years ago
Views:

1 Data Acquisition Ligang Liu Graphics&Geometric Computing Lab USTC

2 Getting Meshes from Real Objects

3 Getting Meshes from Real Objects Many models used in Graphics are obtained from real objects Well known Stanford bunny model?

4 Reverse Engineering Real Object CAD/Graphics Model Build new object

5 Data Acquisition

6 Structure of Data

7 Scan Conversion Main phases Sensing capture raw data Point set Boundary contours Voxels Conversion to polygonal model Followed by Surface reconstruction Mesh improvement Simplification Smoothing

8 Different Acquisition Systems Volumetric scanning Photogrammetry Range scanning

9 2.1 Volume Scanning Build voxel structure by scanning slices

10 Volume Scanning Build voxel structure by scanning slices

11 2.2 Photogrammetry Reconstruction from photographs

12 Photogrammetry Reconstruction from a series of photos (video)

13 2.3 Range Scanning Reconstruction from point cloud Physical real model Acquired point cloud Reconstructed model

14 Range Scanning Systems Passive: stereo matching

15 Range Scanning Systems Active: structured light acquisition

16 Range Scanning Systems Active: laser scanning

17 Examples of Scan Systems

18 Range Scanning Active systems are superior Accurate calibration is crucial Multiple scans required for complex objects scan path planning scan registration Scans are incomplete and noisy model repair, hole filling smoothing for noise removal

19 Range Scanning: Reconstruction Set of raw scans Reconstructed model

20 Sensors Scanner types Laser Imaging (2D/3D) Probing Mixed

21 Sensing Technologies Probing Probing position probe on object record the location Output point cloud data Problematic Labour intensive Error prone

22 Sensing Technologies Imaging Capture multiple 2D images Use image processing tools to create initial geometry data Requirements Many cameras Specific locations

23 3D Imaging Wave based sensors Ultrasound Magnetic Resonance Imaging (MRI) X Ray Computed Tomography (CT) Alternative slice object, take photographs of slices Outputs volumetric data (voxels) contour lines (use imaging techniques)

24 Range Scanners Laser/Optical range scanner provides 2D array of depth data Some capture color (texture) Multiple views for complete object scan: Rotate object Rotate sensor Output point set

25 Model Generation

26 Contours

27 Voxels

28 Triangulating Point Clouds

29 Surface Reconstruction Parametric approaches Neighborhood searching RBF approximation Delaunay triangulation Implicit approaches Distance field Level set Marching cube

30 Limitation Acquire only visible portions Sensitivity to surface properties Confused by interreflections

31 Geometry acquisition Alignment Error controls Properties acquisition Texture Color BRDF BTF Challenges

32 Recap: Digital Media Sound Video Dynamic Geometry Image Geometry

33 Acquisition of Dynamic Objects

34 3D Animation Scanner New technology Record 3D video Active research area Ultimate goal 3D movie making New creative perspectives

35 What is Depth Image (DI)? A pair of aligned maps a texture map I: gives the color of all visible points a depth map D: gives the distance to each visible point

36 Time of flight (ToF) Camera A camera system that creates distance data with help of the time of flight (TOF) principle Light pulses Can measure depth scans at video rate Relatively new devices Become more popular (for everyday users)

37 Range image with height measurements Provide 2.5D structure of the scene

38 Advantages A single depth image Provide 2.5D structure of the scene A set of depth images Might provide hole free 3D scene

39 Set of Depth Images 3D modeling of scene

40 Problems with Depth Images Pros Not only 2D image (2.5D) Easy acquisition Cons Substantial sensor level of random noise low quality data Non trivial systematic bias

41 New Trends: Dynamic Geometry Data Time of flight (ToF) Camera Microsoft Kinects

42 Introduction to Kinect

43 Project Natal Play game without controller Use your body! Proposed in 2009

44 Microsoft Kinect + XBOX 360 Released in Nov. 4, 2010 Cheap: only 150 USD A controller free gaming system Sales 8 million in its first 60 days Breaks Guinness World Records

45 红外发射器彩色 CMOS 红外 CMOS 麦克风马达

46 46

47 成本仅 56 美元

48 原理 2009 年微软收购了以色列 3DV 公司, 让人们以为 Natal 的技术是源自 3DV 的 TOF(time of flight) 2010 年 4 月, 另一家以色列公司 PrimeSense 确认为微软提供了其 light coding 的三维测量技术, 并应用于 Project Natal 不同于 TOF 或者结构光测量技术,light coding 使用连续的照明 ( 而非脉冲 ), 不需要特制的感光芯片, 而只需要普通的 CMOS 感光芯片, 这让方案的成本大大降低 48

49 原理当激光穿透毛玻璃后形成随机衍射斑点, 这些散斑 (laser speckle) 具有高度的随机性, 而且会随着距离的不同变换图案空间中任意两处散斑图案都不同 Light coding 打出了一个具有三维纵深的体编码, 只要看物体表面的散斑图案, 就可以知道这个物体在什么位置 49

50 性能参数水平视角 :57 度垂直视角 :43 度物理倾斜范围 :±27 度景深镜头感应距离 : 米 ( 实测米 ) Z 方向精度 :1 厘米深度图像分辨率 640x480;16 位色深 ;30fps 彩色图像分辨率 640x480;32 位色深 ;30fps 骨骼跟踪系统最多识别 6 人可同时动作捕捉 2 人每个动作捕捉对象上有 20 个捕捉点 50

51 Kinect 破解应用游戏 ( 魔兽体感游戏 ) 人机交互 ( 手势钢琴 ) 机器人监控运动捕捉 ( 测试应用 ) 三维重建 ( 场景人体 ) 51

52 Kinect Data Low resolution 640x480 Very noisy Human skeleton

53 Kinect as Scanners Kinect fusion [Siggraph 2011 Sketch]

54 Scanning Human Bodies [Cui et al., 2011] [Anguelov et al., 2011]

55 Scanning Bodies using Three Kinects [IEEE VR 2012]

56 Kinects are popular More applications Controller everywhere Many opportunities

57 Q&A

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,