Reproduce “Robust Object Detection with Pseudo Labels from VLMs using Per-Object Co-teaching” 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
Uday Bhaskar, Rishabh Bhattacharya, Avinash Patel, Sarthak Khoche, Praveen Anil Kulkarni, Naresh Manwani
Researchers developed a method for training efficient object detectors for autonomous driving by leveraging noisy pseudo-labels from Vision-Language Models (VLMs) and applying a novel per-object co-teaching strategy. This approach resulted in a 15.49 percentage point gain in [email protected] on the KITTI dataset compared to baseline methods, delivering real-time performance suitable for edge devices.
Key insights
- Applying co-teaching at a per-object (anchor-level) granularity is more effective for object detection than image-level filtering, preserving valuable clean labels within partially noisy images.
- The selection loss for filtering noisy labels performs optimally when the objectness term is excluded, as it is an unreliable signal for identifying clean object predictions.
- Employing a ramp-up schedule for the 'forget-rate' in co-teaching provides stable training, allowing models to progressively focus on cleaner examples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.