Reproduce “CS-Net:Contribution-based Sampling Network for Point Cloud Simplification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tian Guo, Chen Chen, Hui Yuan, Xiaolong Mao, Raouf Hamzaoui, Junhui Hou
CS-Net presents a point cloud simplification method that assigns 'contribution scores' to each point, using a differentiable Top-k operation to select the most relevant points for downstream tasks. This approach achieves state-of-the-art performance in tasks such as 3D classification, registration, and compression, while effectively preserving geometric structure and avoiding duplicate points in the sampled output.
Key insights
- Task-aware sampling, guided by learned 'contribution scores' for each point, leads to superior performance in downstream tasks compared to generic simplification methods.
- Formulating the sampling process as a differentiable Top-k operation using optimal transport enables end-to-end training of the network and eliminates duplicate points in the output.
- A joint loss function combining task-specific loss with Earth Mover's Distance loss effectively balances preserving geometric structure and optimizing for task performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.