Adjusting 2D prediction models to be usable for 3D objects
summary
Claims a workable pipeline: starts from STL point‑clouds, uses open3d for centering and ICP + fast global registration to align scans, learns permutation‑invariant point features (PointNet style) and predicts per‑point displacements to forecast how the same object looks five years later. Dataset is paired scans of identical specimens separated by five years. Deliverable is a research prototype that shows plausible forecasts but needs more paired data and GPU compute to be robust.
AI-written summary · about
chapters (27)
- 0:003D object prediction project
- 1:043D prediction challenges
- 2:31Python libraries for 3D
- 3:40STL file format
- 4:23Loading 3D objects in Python
- 5:45Mesh vs point cloud vs voxels
- 7:28Monkey shape prediction use case
- 10:17Aligning 3D objects
- 13:23ICP algorithm
- 16:32Point set registration code
- 18:16Global optimization for alignment
- 19:19Point clouds in 3D modeling
- 20:24Irregular point cloud challenges
- 21:50Permutation invariance solution
- 23:15Feature extraction for point clouds
- 25:25Theory for 3D modeling
- 26:11PointNet architecture
- 27:35Open3D tools
- 28:42Summary and call to action
- 29:04Q&A: 3D printing
- 30:05Q&A: Training data needs
- 30:48Q&A: Project results
- 31:30Q&A: Growth forecasting
- 32:13Q&A: Loading performance
- 32:33Q&A: Graph neural networks
- 33:26Q&A: Training resources
- 34:10Q&A: Prediction output format
related talks
42mDeep Learning approaches meet 3D data
38m3D Face Tracking and Reconstruction using Modern C++
47m3D Digital Avatars with Machine Learning
27mWorking Smart - Machine Learning in Art Asset Creation for the Games Industry
52mMachine learning on point clouds and other varying-size objects
26mStructured Predictions an historical overview
30mConformal Prediction: uncertainty quantification to humanise models
38mActive Learning for 3D Mesh Semantic Segmentation