Reproduce “Black-box Explanation of Object Detectors via Saliency Maps” 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
Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I. Morariu, Ashutosh Mehra, Vicente Ordonez, Kate Saenko
D-RISE, a black-box method developed by a team from Boston University and Adobe Research, generates saliency maps for object detectors to explain both localization and classification decisions. It enables analysis of detector failure modes, reveals reliance on contextual information, and can uncover data biases learned by the models.
Key insights
- Object detectors often learn to utilize contextual information outside an object's immediate boundaries, which D-RISE can reveal through average saliency maps (e.g., a sink's faucet being salient).
- The method can effectively diagnose and visualize subtle, pathological biases learned by detectors from training data, as demonstrated by successfully identifying intentionally introduced synthetic markers.
- D-RISE provides a tool to analyze various failure modes of object detectors, including explaining missed detections, mislocalizations, and misclassifications, by highlighting relevant image regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.