Reproduce “Fast-BEV: Towards Real-time On-vehicle Bird's-Eye View Perception” 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
Bin Huang, Yangguang Li, Enze Xie, Feng Liang, Luya Wang, Mingzhu Shen, Fenggang Liu, Tianqi Wang and 2 more
Fast-BEV presents a framework for real-time Bird's-Eye View perception from multi-camera images, designed for efficient on-vehicle deployment. It achieves a competitive 53.5% NDS on the nuScenes dataset and runs at over 50 FPS, significantly accelerating the view transformation component.
Key insights
- High-performance BEV perception can be achieved with a lightweight, fully convolutional architecture by focusing on efficient view transformation and robust training techniques, avoiding complex depth prediction or transformer models.
- The most significant computational bottleneck in camera-to-BEV projection can be reduced by an order of magnitude on CPU through pre-computation and dense voxel generation, making it feasible for real-time embedded deployment.
- Combining strong 3D data augmentation with multi-frame temporal fusion effectively improves accuracy and robustness, allowing lightweight models to benefit from richer spatial-temporal information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.