https://github.com/nicolas-chaulet/torch-points3d

Unsupversed

Total Denoising: Unsupervised Learning of 3D Point Cloud Cleaning,ICCV19

Task Unsupervised denoising: Our results demonstrate unsupervised denoising performance similar to that of supervised learning with clean data when given enough training examples - whereby we do not need any pairs of noisy and clean training data.

Unsupervised Multi-Task Feature Learning on Point Clouds,ICCV19

Self-supervised Modal and View Invariant Feature Learning,Arxiv2005

PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding,ECCV20,spotlight

Mapping in a Cycle: Sinkhorn Regularized Unsupervised Learning for Point Cloud Shapes,ECCV20

Self-Supervised Learning of Point Clouds via Orientation Estimation,Arxiv2008

Point Cloud Augmentation

PointAugment: an Auto-Augmentation Framework for Point Cloud Classification,CVPR20,oral

Point cloud classification

PointNet++

Point Sampling

Learning to Sample,CVPR19

The input to S-NET is a set of n 3D coordinates, namely points, representing a 3D shape. The output of S-NET is k generated points.

Detail structure of S-Net can be seen in A.1, check code.

G maybe is not a subset P, do matching process by KNN.

construct a sampling regularization loss,composed out of three terms:

\[L_{f}(G,P) = \frac{1}{|G|} \sum_{g \in G} min_{p \in P} ||p - g||^{2}_{2}\] \[L_{b}(G,P) = \frac{1}{|P|} \sum_{p \in P} min_{G \in G} ||p - g||^{2}_{2}\] \[L_{m}(G,P) = max_{g \in G} min_{p \in P} ||g-p||^{2}_{2}\]

\(L_{f}(G,P) + L_{b}(G,P)\) is a typical Chamfer Distance(CD).

SampleNet: Differentiable Point Cloud Sampling,CVPR20

extend previous works by introducing a differentiable relaxation to the matching step, i.e., nearest neighbor selection, during training.

Temperature in softmax is learnable, counter-intuitive.