Weakly-supervised pre-training for 3D human pose estimation via perspective knowledge

Modern deep learning-based 3D pose estimation approaches require plenty of annotations. However, existing datasets lack diversity, which limits the performance current methods and their generalization ability. Although utilize 2D annotations to help estimation, they mainly focus on extracting structural constraints from poses, ignoring information hidden in images. In this paper, we propose a novel method extract weak directly images without supervision. Firstly, perspective prior knowledge generate relative depth human joints. Then, collect dataset (MCPC) labels. Based MCPC, weakly-supervised pre-training (WSP) strategy distinguish relationship between two points an image. WSP enables learning keypoints lots in-the-wild images, is more capable predicting ability for estimation. After fine-tuning model datasets, achieves state-of-the-art results widely-used benchmarks.