D-RISE
A black-box post-hoc explainability method designed for object detection tasks, extending the RISE framework. It generates random occlusion masks on input images, measures the resulting confidence changes, and back-infers the image regions most influential to detection outputs, producing continuous saliency heat maps.
Timeline (last 90 days)
For the D-RISE explanation method, the Spearman correlation coefficient between D-Deletion faithfulness and detection strength falls between 0.70 and 0.82.
The study found that the faithfulness of deletion-based explanations is highly coupled with the detection strength at the time of explanation generation.
After controlling for detection strength within a fixed f0 interval, D-RISE faithfulness still exhibits domain dependence in the intermediate range, with D-Deletion on the PIE dataset being higher than on JAAD, and the difference passing Holm-corrected significance testing.
D-RISE is a post-hoc black-box explanation method designed specifically for object detection tasks, extended from the image classification explanation method RISE; it generates saliency heatmaps by using random occlusion masks to back-infer the image regions with the greatest influence.
EigenCAM's overall faithfulness is lower than D-RISE, but it similarly exhibits score coupling.
All Facts (5)
For the D-RISE explanation method, the Spearman correlation coefficient between D-Deletion faithfulness and detection strength falls between 0.70 and 0.82.
50%UnverifiedThe study found that the faithfulness of deletion-based explanations is highly coupled with the detection strength at the time of explanation generation.
50%UnverifiedAfter controlling for detection strength within a fixed f0 interval, D-RISE faithfulness still exhibits domain dependence in the intermediate range, with D-Deletion on the PIE dataset being higher than on JAAD, and the difference passing Holm-corrected significance testing.
50%UnverifiedD-RISE is a post-hoc black-box explanation method designed specifically for object detection tasks, extended from the image classification explanation method RISE; it generates saliency heatmaps by using random occlusion masks to back-infer the image regions with the greatest influence.
50%UnverifiedEigenCAM's overall faithfulness is lower than D-RISE, but it similarly exhibits score coupling.
50%