Perturbation Detection on COCO (val)
90.9Global AuROC (Untar)Max posterior
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Max posteriorBackbone=Faster R-CNN, Perturbation Strengths=ϵ ∈ {0.25, 0.5, 1, 2, 4, 8, 16}2026.06 | 90.9 | 93.4 | 11.4 | 91 | 49.5 | 44 | 62.5 | |
| HadamardNet (Error)Backbone=Faster R-CNN, Perturbation Strengths=ϵ ∈ {0.25, 0.5, 1, 2, 4, 8, 16}2026.06 | 86.6 | 88.6 | 15.1 | 86.6 | 51.8 | 47.8 | 62.1 | |
| EntropyBackbone=Faster R-CNN, Perturbation Strengths=ϵ ∈ {0.25, 0.5, 1, 2, 4, 8, 16}2026.06 | 84.1 | 87.9 | 14.7 | 84.1 | 50.8 | 46.4 | 60.7 | |
| HadamardNet (Reg.)Backbone=Faster R-CNN, Perturbation Strengths=ϵ ∈ {0.25, 0.5, 1, 2, 4, 8, 16}2026.06 | 72.6 | 75.1 | 70.5 | 72.6 | 50 | 49.1 | 64.2 | |
| HadamardNet (Quantile)Backbone=Faster R-CNN, Perturbation Strengths=ϵ ∈ {0.25, 0.5, 1, 2, 4, 8, 16}2026.06 | 67.6 | 68.1 | 27.6 | 67.8 | 51.7 | 49.3 | 55.1 |