all projects
Point clouds

3D Scene Understanding

From-scratch reimplementations of PointNet, PointNet++, DGCNN and PointMLP for point-cloud classification and part segmentation.

Result DGCNN segmentation: 80.22% mIoU / 91.32% acc · PointNet++: 91.4% acc Stars ★ 7 Tags DGCNN · PointNet++ · PointMLP · ShapeNetPart · ModelNet10
View on GitHub
3D Scene Understanding

A study of the deep-learning toolkit for 3D perception, built by reimplementing the canonical architectures and benchmarking them head-to-head.

Results

  • DGCNN part segmentation on ShapeNetPart: 80.22% mIoU, 91.32% accuracy.
  • PointNet++ classification on ModelNet10: 91.4% accuracy.
  • Clean, comparable implementations of PointNet, PointNet++, DGCNN and PointMLP, so the architectural differences — rather than training tricks — drive the numbers.

The image shows DGCNN segmentation predictions against ground truth: per-point part labels recovered directly from raw 3D coordinates — understanding 3D structure straight from points, with no mesh or voxelization in between.