Reproduce “HGL: Hierarchical Geometry Learning for Test-time Adaptation in 3D Point Cloud Segmentation” on OpenResearch

About this paper

Tianpei Zou, Sanqing Qu, Zhijun Li, Alois Knoll, Lianghua He, Guang Chen, Changjun Jiang

HGL (Hierarchical Geometry Learning) is a framework for Test-Time Adaptation in 3D point cloud segmentation that effectively leverages local, global, and temporal geometric structures for self-supervised learning. The method achieved state-of-the-art performance, with up to a 6.72% mIoU improvement over the source model and an 80% reduction in adaptation time compared to prior approaches like GIPSO.

Key insights

• Comprehensive exploitation of hierarchical geometric information (local, global, temporal) is crucial for robust and accurate self-supervised Test-Time Adaptation in 3D point cloud segmentation.
• Combining local prediction aggregation with geometric purity and object-level prototype learning effectively generates and refines reliable pseudo-labels for adaptation.
• Confidence-weighted temporal consistency regularization stabilizes online adaptation and prevents negative transfer, contributing significantly to both performance and computational efficiency.

Summary by alphaXiv.