Reproduce “SPG: Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point Generation” 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
Qiangeng Xu, Yin Zhou, Weiyue Wang, Charles R. Qi, Dragomir Anguelov
Researchers from Waymo and USC developed SPG, a pre-processing framework that enhances 3D object detection by generating semantic points to recover missing LiDAR data caused by adverse weather. It consistently improves detection performance on both dry-weather (source) and rainy-weather (target) datasets, boosting 3D AP for PointPillars vehicles on Waymo's target domain by 6.7% and achieving state-of-the-art results on the KITTI benchmark.
Key insights
- The primary cause of 3D object detector performance degradation in adverse weather is the substantial deterioration of LiDAR point cloud quality due to missing points.
- Generating semantic points by inferring and adding them to object regions with degraded or missing data can effectively bridge the domain gap caused by adverse weather conditions.
- Training strategies like 'Hide and Predict' and 'Semantic Area Expansion' are crucial for enabling a model to robustly infer and generate semantic points in unobserved yet relevant object areas.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.