Reproduce “Quantifying Data Augmentation for LiDAR based 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
Martin Hahner, Dengxin Dai, Alexander Liniger, Luc Van Gool
Researchers from ETH Zurich and Toyota TRACE teams conducted a systematic, quantitative study of data augmentation techniques for LiDAR-based 3D object detection, revealing significant individual contributions and non-intuitive effects while proposing an improved augmentation policy. The proposed policy boosted 3D mAP40 on moderate cars by 1.5% on KITTI and generalized across multiple state-of-the-art models and datasets.
Key insights
- Data augmentation is critically important, accounting for nearly 23% of a model's overall performance in LiDAR-based 3D object detection.
- Global augmentations like rotation, translation, and scaling provide substantial performance gains, while local translation can be counterproductive and ground removal is surprisingly detrimental.
- Filtering out 'hard' examples during training can counter-intuitively improve performance on those same 'hard' cases during evaluation, suggesting potential issues with model learning from overly difficult instances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.