Knowledge distillation for BEVFusion
BEVFusion merges camera images and LiDAR into one bird's-eye-view map to find cars, people and obstacles in 3D. It is accurate and heavy. I distilled it into a compact student built from YOLO and PointPillars, and benchmarked it on nuScenes.
- Student on its own
- Class weights for rare objects
- Learning from the teacher
- Both combined
- PyTorch
- nuScenes
- Docker
- CUDA
- RTX 3090