Pose-to-Pose: A New Task and Benchmark for Human Pose Transition in Yoga

Bhat Dittakavi, Swarnim Maheshwari, Vineeth N. Balasubramanian; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 6091-6100

Humanposetransition(HPT)involvestransformingtheskeletal representation of a person from a source pose to that of the same person in a target pose. This new task is crucial for various applications, such as personalized sports training, animation, and motion capture. Despite its importance, there is little to no existing work in this emerging area. To address this gap, we introduce a new benchmark with our novel method, Pose-to-Pose. Pose-to-Pose integrates algorithmic recourse with personalized anatomical constraints, uses exemplars to capture target pose characteristics, and leverages counterfactuals from a pre-trained pose classifier. Wealso present PoseHPT, a new Yoga dataset featuring pairs of source and target poses. Experimental results demonstrate that Pose-to-Pose outperforms its baselines. To evaluate how well the method generalizes across different source poses to produce a consistent target pose, we introduce the Target Pose Consistency Metric (TPCM). To assess the accuracy of aligning the target pose with the exemplar and ensure precise representation of pose characteristics, we introduce the Class ExemplarAlignmentScore(CEAS).Ourcontributions--Poseto-Pose, the new dataset, and the metrics--pave the way for further research in realistic human pose transition.