Deep Learning approaches meet 3D data
summary
Covers concrete methods: PointNet with MLP + max‑pool and T‑net, and PointNet++ with sampling + grouping for local context; shows segmentation and classification results and which critical points matter. Describes mesh reconstruction attempts (ball pivoting, Poisson), failure modes on noisy real scans, and Laplacian spectral vs geodesic/spatial convolutions with functional maps for correspondences. Mentions datasets (ModelNet/PartNet/ABC), Python/C++ libraries (Open3D, Kaolin, libIGL) and a robotics project that uses point‑cloud segmentation and morphing — results are promising but reconstruction has issues.
AI-written summary · about
chapters (10)
- 0:00Start
- 0:19Introduction of 3d Data
- 1:52How 3d Data Can Be Represented
- 2:32Difference between 2d and 3d for Computer Vision
- 4:31Pointnet
- 5:09Feature Transformer Block
- 7:34Learn Patterns of the Network
- 8:31Semantic Segmentation Results
- 11:14Spectral Domain
- 37:48Are There any Solutions for Using Sequential Frames from a Single Camera for 3d Reconstruction
related talks
38mActive Learning for 3D Mesh Semantic Segmentation
35mAdjusting 2D prediction models to be usable for 3D objects
25mDeep Learning in Biomedicine: Challenges and Opportunities
01h 30mDeep Learning Techniques for Monocular Depth Estimation
41mDeep Learning to Solve Challenging Problems
19mDeep Learning is Just LEGO: & Other Hands on Machine Learning Activities
36mDeep Learning - Segmentation, Identification and Classification
54mIntroduction to Understanding Deep Learning