Reproduce “Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming” 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

Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S. Ecker, Matthias Bethge, Wieland Brendel

Researchers from the University of Tübingen established the Robust Detection Benchmark, comprising PASCAL-C, COCO-C, and Cityscapes-C, to systematically assess object detection model robustness against 15 image corruption types. They found that training models with a combination of clean and stylized data consistently improves robustness under synthetic and natural adverse conditions while largely preserving clean data performance.

Key insights

  • State-of-the-art object detection models show substantial performance drops, with some losing over two-thirds of their clean performance on corrupted data, regardless of backbone capacity or head architecture complexity.
  • Data augmentation via style transfer is an effective and low-cost method to enhance object detection robustness against diverse corruptions and natural distortions without requiring architectural changes or additional human labeling.
  • Improved performance on the synthetic Robust Detection Benchmark correlates with increased robustness against real-world natural adverse conditions such as fog, rain, snow, and nighttime.

Summary by alphaXiv.

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