Reproduce “MoDAR: Using Motion Forecasting for 3D Object Detection in Point Cloud Sequences” 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

Yingwei Li, Charles R. Qi, Yin Zhou, Chenxi Liu, Dragomir Anguelov

MoDAR enhances 3D object detection by using motion forecasting as a virtual sensor modality to integrate temporal information from point cloud sequences. This method improves pedestrian detection by 13.1 mAPH and vehicle detection by 9.0 mAPH on the Waymo Open Dataset, effectively scaling to long sequences without high computational costs.

Key insights

  • Treating motion forecasting as a virtual sensor modality provides a lightweight and scalable way to integrate long-term temporal information into 3D object detection.
  • Inverting the traditional perception pipeline by using motion forecasting to enhance detection, rather than vice versa, can improve detection quality, especially for occluded or distant objects.
  • Early fusion of MoDAR points with raw LiDAR data, allowing the detector to learn feature combinations, consistently outperforms simpler late fusion strategies.

Summary by alphaXiv.

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