Human Pose Estimation via Multi-resolution Convolutional Neural Network

Human pose estimation is a challenging problem in computer vision tasks and shares all the difficulties of object detection. This paper focuses on the problems of estimating human pose in still images, including the various appearances and non-rigid body parts. To address these problems, we adopt a CNN to extract multi-scale part information. The appearance model is learned by CNN and the deformable model is computed based on appearance feature. Then, a Multi-Resolution Convolutional Neural Network (MR-CNN) is proposed to train and learn the multi-scale feature of each body part. This model is compared with the related work on the the Leeds Sport Dataset (LSP). The experimental results demonstrate the effectiveness of the proposed method.