Human arm pose modeling with learned features using joint convolutional neural network

This paper proposes a new approach to model human arm pose configuration from still images based on learned features and arm part structure constraints. The subjects in still images have no assumption with regards to clothing style, action category and background, so our model has to accommodate these uncertainties. Proposed approach uses an energy model that incorporates the dependence relationships among arm joints and arm parts, where the potentials represent their occurrence probabilities. Positive and negative instances are computed from input image, using multi-scale image patches to capture the details of arm joints and arm parts. A joint convolutional neural network is then developed for feature extraction. Local rigidity of arm part is used to constrain occurrence of arm joints and arm parts, and these constraints can be efficiently incorporated in dynamic programming for human arm pose inference. Our experimental results show better performance than alternative approaches using hand-crafted features for various still images.