Reproduce “OT-Drive: Out-of-Distribution Off-Road Traversable Area Segmentation via Optimal Transport” 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

Zhihua Zhao, Guoqiang Li, Chen Min, Kangping Lu

OT-Drive presents a multi-modal fusion framework for off-road traversable area segmentation that leverages optimal transport and scene-conditioned distribution alignment to improve out-of-distribution (OOD) generalization. The method achieved 95.16% mIoU in OOD scenarios on the ORFD dataset, outperforming baselines by 6.35% and exhibiting minimal performance degradation between known and unknown environments.

Key insights

  • Modeling complex scene variations through marginal probabilities of environmental attributes allows for compositional generalization to novel combinations not explicitly seen during training.
  • Optimal transport provides a flexible mechanism for robust, distribution-level feature alignment, overcoming the limitations of pixel- or token-level fusion methods under OOD conditions.
  • Leveraging Vision-Language Models (VLMs) to generate scene anchors offers domain-invariant semantic guidance, grounding visual features in abstract linguistic concepts for improved robustness.

Summary by alphaXiv.

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