Course video link: GAMES 203 course: "3D Reconstruction and Understanding"
Lecturer: Huang Qixing, University of Texas at Austin
Course homepage: https://www.cs.utexas.edu/~huangqx/Games_3D_Recons_Understanding.html
Recommended bibliography: " An Invitation to 3-D Vision》Yi Ma /《Point-based Graphics》/《Polygon Mesh Processing》
For more learning materials related to computer graphics, please visit:
Computer Graphics and Mixed Reality Online Platform GAMES: http://games -cn.org
P1 Introduction and (3D) Scanning
1.1 Introduction
- 3D Vision: Pose Estimation, SfM, Multi-view Stereo
- understanding: Classification, Segmentation, Detection
- Some large-scale online datasets : 3D Warehouse, Yobi3D, 3DFront
- The difference between 2D and 3D : There are many different representation methods for 3D data, such as voxel, grid, point cloud, multi-view, scene graph, semantic segmentation, etc.
- 应用:3D mapping,Performance capture, Robotics, Autonomous Driving, Reverse Engineering
- Topics:
(1) 3D Reconstruction : scanning, scan registration, surface reconstruction, SfM, Multi-view stereo, map synchronization (linked to graph-based SLAM) (2)
How to represent 3D Data , Conversion between different representations
(3) How to understand 3D Data : Matching, Retrieval, Segmentation, Classification&Clustering - Numerical Optimization is important, there are plenty of optimization softwares.
- Categories of optimization models: linear/nonlinear, convex/nonconvex, continuous/discrete, deterministic/with uncertainty
- Optimization Examples:Bundle Adjustment,Surface Fitting,MRF Inference
- Homework: (1) Dense reconstruction; (2) Primitive Extraction.
- Possible project topics: (1) image-based modeling from internet images ; (2) single-view reconstruction from some real images . (3) Neural networks for 3D representation ; (4) Geometry understanding ; (5) Reconstruction based on hybrid sensors . (6) Robotic 3D vision (prominent perspective, NBV, robot grasping, human-object interaction, model-based viewpoint planning, active vision, exploration strategy); (7) Structural restoration through optimization (human pose) (8) Map Synchronization; (9) Uncertainty;
48 min
1.1 Scanning
- Geometry Reconstruction Pipeline
- Depth Sensing: (1) Contact; (2) Transmissive; (3) Reflective;
- Optical methods: (1) Passive (from image); (2) Active (Stereo, ToF, Triangulation)
- Pattern Design …
P2 Registration (registration, integration information)
P3 surface reconstruction
P4 Structure From Motion
P5 Multi-View Stereo
Map synchronization in P6 Inverse Problems
P7 point cloud processing
P8 Mesh Processing
P9 3D Deep Learning
AlexNet brings mutations to deep learning research for computer vision.