Which paper are you mainly refered?

[MRJTM]

I am new in the field of ReID and I want to run a project that implement ReID with triple loss, it would be better for me if the project is based on a specific paper. So I would appreciate that if you could tell me if you had mainly refered a paper and which paper it is.

[layumi]

You may refer to this paper.
https://arxiv.org/abs/1703.07737
But I change the learning policy and sampler method a little.

[MRJTM]

Thank you for replying me，I‘ll read this paper then run your code to understand how you train a model with TriHard loss.

[upgirlnana]

i can’t understand the positive and negative samples' choosing process,can you describ it more detail

[MRJTM]

For triplet loss， In his code, you can see that he first random select K samples as anchors, K is the batch size, and then for each id in these K sample, he select N samples to generate positive and negative sample,
so you'll get K ids, for each id ,you have 1+N samples, the 1 is the anchor, and you can get the hardest positive sample from these N samples, for hardest negative sample , it is chosen from the left (K-1)N negative samples. So finally you will get K triplets:(anchor, hardest positive, hardest negative)

[minygd]

For triplet loss， In his code, you can see that he first random select K samples as anchors, K is the batch size, and then for each id in these K sample, he select N samples to generate positive and negative sample,
so you'll get K ids, for each id ,you have 1+N samples, the 1 is the anchor, and you can get the hardest positive sample from these N samples, for hardest negative sample , it is chosen from the left (K-1)N negative samples. So finally you will get K triplets:(anchor, hardest positive, hardest negative)

Thx!

[wudibaichi2]

Can you tell us more about your sampling strategy? I don't understand it well?