Reproduce “Motion Inspired Unsupervised Perception and Prediction in Autonomous Driving” 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
Mahyar Najibi, Jingwei Ji, Yin Zhou, Charles R. Qi, Xinchen Yan, Scott Ettinger, Dragomir Anguelov
Waymo LLC introduces a motion-inspired unsupervised framework for autonomous driving, enabling perception and prediction of open-set moving objects without human labels. This approach generates high-quality pseudo-labels, achieving competitive 3D detection performance and substantially improving trajectory prediction for previously unknown categories on the Waymo Open Dataset.
Key insights
- NSFP++ significantly advances unsupervised scene flow estimation by employing robust static point identification, a divide-and-conquer strategy via scene decomposition, and local flow consistency regularization within object components.
- The AML pipeline effectively generates object proposals through spatio-temporal clustering of points and flows, followed by multi-object tracking and motion-inspired sequential ICP for amodal bounding box refinement.
- Motion cues alone, extracted without semantic labels, are sufficient to identify, track, and generate high-fidelity perception and prediction labels for arbitrary moving objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.