Reproduce “LiDAR Distillation: Bridging the Beam-Induced Domain Gap for 3D Object Detection” 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
Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, Jiwen Lu
The LiDAR Distillation framework facilitates effective 3D object detection across varying LiDAR beam configurations by bridging the beam-induced domain gap through progressive knowledge transfer and density alignment. This approach allowed a model finetuned with just 5% of target domain labels to surpass one trained from scratch on 100% of labels, achieving 47.16 APBEV compared to 45.31 APBEV.
Key insights
- Aligning the point cloud density between the source (high-beam) and target (low-beam) domains is critical for effective adaptation of 3D object detectors, as detectors are highly sensitive to density variations.
- Downsampling high-beam source data to create pseudo low-beam data is more effective than attempting to upsample low-beam target data, as upsampling can introduce noisy points and degrade object features.
- Knowledge distillation, particularly by focusing on informative ROI regions within BEV feature maps, successfully transfers rich information from high-resolution data to density-aligned low-resolution data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.