Reproduce “DriveFlow: Rectified Flow Adaptation for Robust 3D Object Detection 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

Hongbin Lin, Yiming Yang, Chaoda Zheng, Yifan Zhang, Shuaicheng Niu, Zilu Guo, Yafeng Li, Gui Gui and 2 more

DriveFlow employs a rectified flow adaptation technique to create diverse training data for resilient 3D object detection in autonomous driving. This approach adapts pre-trained text-to-image flow models to maintain 3D object geometry, achieving augmented image generation over 30 times faster than previous methods while improving detection accuracy in out-of-distribution conditions.

Key insights

  • Decomposing velocity fields into low-frequency and high-frequency components allows for independent and targeted manipulation of scene context and foreground objects.
  • High-frequency information within object regions is essential for preserving accurate 3D geometry, which enables the direct reuse of existing bounding box annotations.
  • Adapting the velocity field of pre-trained T2I flow models for specific editing tasks, rather than training from scratch or using inversion, offers a highly efficient and controllable augmentation strategy.

Summary by alphaXiv.

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