Towards Locality Similarity Preserving to 3D Human Pose Estimation

Estimating 3D human pose from an annotated or detected 2D pose in a single RGB image is a challenging problem. A successful way to address this problem is the example-based approach. The existing example-based approaches often calculate a global pose error to search a single match 3D pose from the source library. This way fails to capture the local deformations of human pose and highly dependent on a large training set. To alleviate these issues, we propose a simple example-based approach with locality similarity preserving to estimate 3D human pose. Specifically, first of all, we split an annotated or detected 2D pose into 2D body parts with kinematic priors. Then, to recover the 3D pose from these 2D body parts, we recombine a 3D pose by using 3D body parts that are split from the 3D pose candidates. Note that joints in the combined 3D parts are refined by a weighted searching strategy during the inference. Moreover, to increase the search speed, we propose a candidate selecting mechanism to narrow the original source data. We evaluate our approach on three well-design benchmarks, including Human3.6M, HumanEva-I, and MPII. The extensive experimental results show the effectiveness of our approach. Specifically, our approach achieves better performance than compared approaches while using fewer training samples