Visual Computing for Industry, Biomedicine, and Art

Table 1 Perturbed performance comparisons of different attack types on the COCO validation set for HigherHRNet

From: Local imperceptible adversarial attacks against human pose estimation networks

	AP	AP⁵⁰	AP⁷⁵	AP^M	AP^L	AR	AR⁵⁰	AR⁷⁵	AR^M	AR^L
Clean	0.671	0.862	0.730	0.615	0.761	0.718	0.885	0.768	0.651	0.814
BIM	0.499	0.706	0.538	0.411	0.627	0.543	0.723	0.579	0.441	0.684
PGD	0.463	0.723	0.494	0.395	0.568	0.541	0.770	0.573	0.454	0.661
C&W	0.651	0.842	0.710	0.578	0.760	0.695	0.865	0.742	0.610	0.815
Ours	0.605	0.802	0.658	0.516	0.739	0.653	0.826	0.696	0.552	0.796

AP⁵⁰ is the AP at IOU = 0.5, AP⁷⁵ is the AP at IOU = 0.75, AP^M is the AP for medium objects, AP^L is the AP for large objects

Back to article page