Recurrent Transformer Networks for Semantic Correspondence

first paper I’ve read on semantic correspondence

• objective: a field from src img to tgt img
• previous approaches
  1. infer the field in feature extraction step, by min. a objective function
  2. feature extraction -> predict field in a end2end manner
• this paper
  • conbine these two
    • feature extraction -> predict a field -> modify the feature using the field -> …
  • matching phase: similarity using cosine distance
  • loss: “classification”, does that mean this transformation should be better than other ones?