IIPA-Net: Joint Illumination-Invariant and Pose-Aligned Feature Learning for Person Reidentification

Algorithm 1

Training steps of the proposed network.

  1. The shared-weights Resnet-50 is trained to convergence with triplet loss.

  2. All synthetic images, together with their original images, are fed into the Retinex decomposition network.

  3. Parts attention module is trained using the training images set.

  4. The whole network is fine-tuned with Equation (10).