Reproduce “Benchmarking and Improving Bird's Eye View Perception Robustness in Autonomous Driving” 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
Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, Ziwei Liu
Researchers from several institutions including the University of California, Irvine, and Shanghai AI Laboratory, introduced RoboBEV, a benchmark for evaluating the resilience of bird's eye view (BEV) perception models in autonomous driving to natural corruptions and sensor failures. The study found all benchmarked models experience performance degradation under these conditions, and a two-stage fine-tuning strategy leveraging CLIP as a backbone improved NDS scores for Dark, Fog, and Snow corruptions by up to 23.1%.
Key insights
- Depth-free BEV transformation methods generally exhibit superior robustness compared to depth-based approaches, which show severe degradation due to errors in intermediate depth estimation under corruptions.
- Model pre-training and the effective utilization of temporal fusion strategies significantly enhance robustness, particularly against semantic and temporal corruptions, respectively.
- Foundation models, such as CLIP, can effectively transfer their generalization capabilities to BEV perception tasks, with a two-stage fine-tuning strategy demonstrating notable improvements in robustness under various corruptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.