Week 08 — Point Cloud Representations

Three architectures for unordered sets of 3D points: PointNet, PointNet++, and EquivariantConv.

Lecture

Point clouds as permutation-invariant inputs · PointNet (Qi et al. 2017) and its critical-point analysis · PointNet++ as hierarchical PointNet · DGCNN (Wang et al. 2019) · current LiDAR / point-cloud benchmarks.

Read before the lecture

• Qi et al., *PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation* (CVPR 2017)

Code lab

PointNet on ModelNet10

Train a PointNet for 3D shape classification on ModelNet10. Evaluate equivariance breakage by rotating test inputs.

Notebook: lab03-pointnet.ipynb  ·  Dataset: ModelNet10 (Princeton).

Reference text for this week: chapter 08 of the bilingual notes — EN PDF · FR PDF.