Reproduce “EvObj: Learning Evolving Object-centric Representations for 3D Instance Segmentation without Scene Supervision” 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
Jiahao Chen, Zihui Zhang, Yafei Yang, Jinxi Li, Shenxing Wei, Zhixuan Sun, Bo Yang
EvObj introduces a framework for unsupervised 3D instance segmentation without scene supervision, effectively bridging the geometric domain gap between synthetic pretraining and real-world point clouds. The method achieves superior performance on ScanNet and S3DIS datasets by incorporating evolving object-centric representations to handle morphological variations, partial geometries, and noise.
Key insights
- Continuously evolving the object discerning module using self-supervised pseudo-labels allows the system to dynamically adapt its object priors to real-world morphological variations and context.
- Reconstructing complete object shapes from partial scans significantly enhances the objectness assessment of occluded instances, which are prevalent in real-world 3D data.
- Integrating these specialized modules allows for iterative refinement of object candidates and their assessment, effectively mitigating the synthetic-to-real geometric domain gap for unsupervised 3D instance segmentation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.