Fig. 2
From: Local imperceptible adversarial attacks against human pose estimation networks
Comparing perturbation crafted by PGD, C&W and ours in various scenarios on HigherHRNet. Adversarial examples of our method are locally perturbed and able to truly maintain imperceptibility. In all cases, PGD and ours can fool HigherHRNet to predict incorrect poses or even to fail in detecting human joints, while C&W cannot always succeed. First column: human poses predicted on clean data. Second column: human poses attacked by PGD. Third column: human poses attacked by C&W. Fourth column: human poses attacked by our method