Reproduce “SUG-Occ: Explicit Semantics and Uncertainty Guided Sparse Learning for Efficient 3D Occupancy Prediction” 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

Hanlin Wu, Pengfei Lin, Ehsan Javanmardi, Naren Bao, Bo Qian, Hao Si, Manabu Tsukada

A framework for efficient 3D semantic occupancy prediction, SUGOcc, from The University of Tokyo utilizes explicit semantic and uncertainty guidance to learn sparse representations for autonomous driving, achieving a 14.91% mIoU on SemanticKITTI and 37.01% RayIoU on Occ3D-Nuscenes at over 10 FPS.

Key insights

  • Incorporating explicit semantic and depth uncertainty priors during 2D-to-3D lifting enables the selective projection of only informative features, reducing redundancy and improving sparse representation quality.
  • Explicit unsigned distance encoding can preserve geometric consistency along camera rays, preventing blurring that occurs with probabilistic feature aggregation.
  • A cascade sparse completion network with novel hyper cross convolutions and adaptive pruning is effective for efficient coarse-to-fine reconstruction, while an OCR-based mask decoder can refine semantic predictions through lightweight query-context interactions.

Summary by alphaXiv.

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