An Improved Deep Learning Architecture for Person Re-Identification

Posted on 14/04/2019, in Paper.
  • Overview:This paper proposes(in the first time?) an end2end feature extracting/comparison system and achieves STOA. It proposes a new layer to generate a similarity score from the feature maps.
  • Tied-conv: Tied conv for feature extraction.
  • Cross-input neighbor difference: For a given position, it will copy image A’s pixel to 5 by 5 region and subtract the 5*5 neighborhood of image B element-wisely. This will be done in a reverse way as well.
  • Hard Negative Mining: Downsample the negative samples, train the model, predict the negative and refine the last layer with the worst examples.