Adaptive Anchor for Robust Keypoint Localization

Dongkai Wang, Shiliang Zhang

Keywords: keypoint localization, human pose estimation

Abstract: Existing keypoint localization methods mostly select pre-defined points like image center as anchors, then infer keypoint locations referring to anchors. Pre-defined anchors are sensitive to occlusions and crowded scenes, leading to degraded robustness. This paper proposes to detect Adaptive Anchor (AdaAnchor) for keypoint localization. Instead of relying on pre-defined rules, AdaAnchor is adaptively selected by maximizing both the keypoint localization confidence and accuracy. This strategy leads to more robust keypoint localization even with the existence of occlusions and truncations. AdaAnchor can be flexibly integrated into different methods by replacing their anchor point selection strategies. Experiments show that it surpasses previous anchor selection methods on both single and multiple keypoint localization tasks. For instance, replacing the heatmap-anchor with AdaAnchor reduces the localization error of invisible keypoints by 6%, meanwhile improves the confidence by 41.7+% on COCO in single keypoint localization. This advantage sustains on multiple keypoint localization task, e.g., AdaAnchor outperforms heatmap-anchor by 4.8% AP on bottom-up multi-person pose estimation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)