Reproduce “Point-GCC: Universal Self-supervised 3D Scene Pre-training via Geometry-Color Contrast” 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
Guofan Fan, Zekun Qi, Wenkai Shi, Kaisheng Ma
Point-GCC, a self-supervised pre-training framework from Xi an Jiaotong and Tsinghua Universities, explicitly aligns 3D geometry and color information through hierarchical contrastive and reconstructive learning. This approach achieved an 18.3% mIoU in fully unsupervised semantic segmentation on ScanNetV2 and consistently improved 3D object detection and instance segmentation across datasets, setting new state-of-the-art results for object detection on SUN RGB-D and S3DIS.
Key insights
- Explicitly contrasting and reconstructing geometry and color information, rather than treating them uniformly, yields richer and more discriminative 3D representations.
- Integrating both point-level and object-level supervision effectively bridges the gap between pre-training tasks and the object-centric nature of downstream 3D scene understanding tasks.
- The architecture-agnostic design ensures the framework's broad applicability, allowing it to enhance performance across diverse 3D backbones (e.g., PointNet++, MinkResNet) and task heads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.