Reproduce “The Elephant in the Room” 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
Amir Rosenfeld, Richard Zemel, John K. Tsotsos
Researchers from York University and the University of Toronto investigated robustness issues in state-of-the-art object detectors by transplanting objects into new image contexts. They found that these insertions frequently caused non-local detection instabilities, misclassifications, and undetected objects across various models, even when the transplanted object did not occlude existing detections.
Key insights
- Object detectors are highly sensitive to contextual changes, exhibiting non-local effects where a transplanted object alters the detection of non-overlapping objects in the scene.
- Feature interference from pixels both inside and outside an object's precise bounding box contributes to its classification and detection outcomes.
- The observed failures are not solely due to out-of-distribution object pairings but also stem from limitations in handling local context, partial occlusions, and potentially artifacts of non-maximal suppression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.