Multi-Branch Convolutional Auto-Encoder Model For Cross-Domain Person Re-Recognition

Cross-domain person re-recognition is very important for the applications of intelligent video surveillance. To further reduce the cross-domain differences, a new multi-branch convolutional auto-encoder based cross-domain person re-recognition network (MCAENet) is here proposed. Firstly, the reconstructed feature map is utilized as an input of the generative model module, and reconstruction loss and divergence loss are utilized to extract domain specific features and unknown domain features; then, the domain invariant feature guidance branch is introduced to enhance the robustness of domain invariant features; finally, identity loss is introduced and combined with the triple loss to constrain the distance between different pedestrians, so as to extract representative features to distinguish different pedestrians. Extensive experimental results demonstrate that the proposed cross-domain person re-recognition method based on convolutional -encoder achieves the best identification accuracy in various environments.