Reproduce “Training Region-based Object Detectors with Online Hard Example Mining” 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

Abhinav Shrivastava, Abhinav Gupta, Ross Girshick

Online Hard Example Mining (OHEM) introduces a method to integrate principled hard example selection directly into the online training of region-based object detectors. This approach improves detection performance, yielding up to 6.6 mAP points higher on MS COCO and achieving new state-of-the-art results on PASCAL VOC when combined with other techniques.

Key insights

  • Principled hard example mining can be effectively re-integrated into online SGD training, overcoming the limitations of prior heuristic approaches in modern deep object detectors.
  • Dynamically selecting training examples based on the current loss allows the model to continuously focus its learning efforts on the most challenging aspects of the data.
  • An efficient architecture employing separate "readonly" and "regular" RoI networks enables practical implementation, performing a full forward pass but a backward pass only for selected hard examples.

Summary by alphaXiv.

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