Overview: This paper advocates the triplet loss by proposing batch-hard triplet loss, combine it with softmax and systematically evaluate the performance of framing ReId problems.
Batch-hard loss: Instead of randomly sampling all positive/negative pairs, the author suggests we can focus on the hardest positive sample pair and negative sample pair constructed by a batch. Comparing with other triplet loss this works the best in ReID.
I, V and E: Models for ReID are classified into three framings: Identification - frame the problem as classification; Verification - frame the problem as two image inputs, outputting if they are from same identify; and Embedding - Pulling for the same class and pushing for different classes in the embedding space w.r.t to some distance measure. E usually performs the best.
Pre-trained: The author also argued given enough data, training from scratch accommodate the specific needs of ReIDs better. And triplet loss could help with training from scratch.
Todo:there are still some items I am not sure I understand in this particular such as, Lifted embedding loss, offline hard-mining, ReRanking.