POSE-AWARE PERSON RECOGNITION

Abstract

This project develops pose representation for the person's recognition. Although objects like face are easy to align, but it is more difficult to align the parts of the human body with big changes. As a result, this project presents a specific visual model that receives the recognition while keeping the human body in mind. This project divides the place of human pose into limited groups, in which each sample is included in special orientation or body perspective. Unlike other project this project, learn the representation of multi-sectoral infectious convolutional neuronal network (CNN) for each point. However, unlike previous approaches to creating a neural network for each area ofthe body, we jointly optimized the network in many areas of the body. It provides additional flexibility for the network to create predictions based on some informative body areas. It contradicts with a different training that strictly applies the correct predictions of each area of the body during testing. We obtain samples of identifying predictions through linear combination of classified scores, each of which is formed by the specific representation of the pose. To combine classifiers, the weight is obtained through a pose estimator, which calculates the probability of each view.