PROJECT TITLE :

Deep Representation Learning With Part Loss for Person Re-Identification

ABSTRACT:

It is essential for a person's re-identification to learn discriminative representations of unseen images (ReID). Deep representations are typically learned in classification tasks to reduce the probability of incorrect classifications on the training set. The training set for a discriminative human body component was over-fitted in our studies, as we discovered. Part loss network, a deep representation learning process, is proposed to reduce both the risk of incorrectly classifying training person photos and the risk of incorrectly classifying unseen person images. The danger of representation learning is assessed by the proposed part loss, which automatically detects human body parts and computes the person classification loss for each part individually.. It is more effective to consider the loss of specific body parts in addition to the loss of global classification when training the deep network for discrimination of unseen individuals. Our representation outperforms previous deep representations on three person ReID datasets, namely Market1501, CUHK03, and VIPeR.