Deep self-paced learning for person re-identification

Highlights：

• We propose a novel deep self-paced learning algorithm to supervise the learning of deep neural network, in which a soft polynomial regularizer term is proposed to gradually involve the faithful samples into training process in a self-paced manner.

• We optimize the gradient back-propagation of relative distance metric by introducing a symmetric regularizer term, which can convert the back-propagation from the asymmetric mode to a symmetric one.

• We build an effective part-based deep neural network, in which features of different body parts are first discriminately learned in the convolutional layers and then fused in the fully connected layers.