Reproduce “Fast Occupancy Network” 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

Mingjie Lu, Yuanxian Huang, Ji Liu, Xingliang Huang, Dong Li, Jinzhang Peng, Lu Tian, Emad Barsoum

Researchers from AMD developed the "Fast Occupancy Network," an efficient vision-based model for 3D semantic occupancy prediction in autonomous driving. This model delivers higher accuracy than previous state-of-the-art while being up to three times faster on AMD MI250 GPUs.

Key insights

  • 2D deformable convolutions can effectively and efficiently lift 2D BEV features to a 3D voxel representation, demonstrating superior performance over simpler methods and far greater efficiency than 3D deformable attention.
  • Significant computational savings in 3D multi-scale feature fusion can be achieved by predominantly applying 2D convolutions across height layers, with minimal, targeted 3D convolutions at coarser resolutions.
  • Incorporating "cost-free" auxiliary supervision from perspective view segmentation during training, even with sparse ground truth, effectively strengthens the image backbone and improves overall occupancy prediction accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.