Deep Active Learning For Human Pose Estimation Via Consistency Weighted Core-Set Approach

Aiming to develop an annotation-efficient algorithm by selecting the most informative samples to be labeled for human pose estimation, we propose a novel active learning approach that generalized the standard core-set approach by dynamically incorporating the uncertainty and representativeness cues. Based on our designed assignment cost that consists of fast spatial consistency and domain-shift scaled distance, our method proposed an adaptive information measurement as data selection criterion. Extensive experiments conducted on two network architectures and various datasets demonstrate the effectiveness and superiority of our proposed approach in comparison with other state-of-the-art methods.