Human outline keypoints detecting via global and grouping strategy.

Different from human's pose estimation, the outline keypoints detecting task has not yet long been researched sufficiently in computer vision field. Body's outline cannot be directly recovered with joint keypoints or skeleton only, even with the aid of semantic segmentation. Detecting points of human's outline is still a challenging and relatively new work which aims at describing the outline shape of a human being with ordered keypoints. Moreover, the estimation must be robust with interference, such as self-occlusion or complicated background. By analyzing the characters of the task, we put forward global and grouping strategy. Based on this, we introduce a method to regress 63 keypoints in real-time with outstanding capability even in mobile device. Experimental results show that the proposed model has excellent state-of-the-art performance over traditional pose estimation models.