Reweighted sparse representation with residual compensation for 3D human pose estimation from a single RGB image.

3D human pose estimation from 2D joints of an image is a worthwhile and challenging research topic. Since a specific 2D pose could be projected from various 3D poses, the ambiguity becomes a difficult obstacle when recovering 3D pose from 2D. Many supervised learning solutions have been proposed in recent years, however, most of them require an abundant of well-annotated training samples to get satisfied estimation performance. In this paper, an unsupervised approach that built upon sparse representation (SR) is presented and its two enhancement schemes are provided. As the first scheme, the reweighting one improves the sparsity of the SR model which leads to a more accurate solution. Furthermore, based on the resulting minimization residual of the loss function in 2D, the discrepancy between the estimated 3D pose and the target 3D pose is used to adjust the estimated 3D pose to reach a better accuracy. Comprehensive experiments have been conducted on four well-recognized benchmarks for evaluation. Significant and consistent improvements over existing SR models are observed in our experiments. Furthermore, the proposed approach even outperforms many supervised learning works.