Human Pose Estimation with Deeply Learned Multi-Scale Compositional Models

Compositional models are meant for human pose estimation (HPE) due to their abilities to capture relationships among human body parts. Deeply learned compositional model (DLCM) utilizes deep neural networks to learn compositionality of human body parts and has achieved great improvements in human pose estimation. The DLCM has a hierarchical compositional architecture and bottom-up/top-down inference stages. The previous works have proven that multi-scale deep features are beneficial for computer vision tasks, such as classification and human body keypoints detection. However, learning multi-scale feature pyramids in DLCM has not been well explored. In this paper, we propose a new method to apply the multi-scale feature pyramid module to further improve the performance of the DLCM, which is named as deeply learned multi-scale compositional model (DLMSCM). We design multi-scale residual modules as the basic blocks to learn multi-scale deep features which can capture the scale variations of different body parts. With the multi-scale mechanism in the framework of the DLCM, the model can not only deal with scale variations of body parts but also find joints dependencies, therefore enforce the entire body joints structural constrains. As a result, more precise body keypoints detection can be acquired. Our approach outperforms the other state-of-the-art methods on two standard benchmarks datasets MPII and LSP for human pose estimation.