Some queries regarding the model design about person-reid-triplet-loss-baseline

[huanghoujing]

Hi, thanks for your attention.

1.1. No other changes are made.

1.2. I find that vanilla average pooling is really well, so I did not add additional embedding layer after ResNet-50 in the released code. I also did experiments with one additional embedding layer, the result of which I can not remember now. So I also did not provide results for it.

1.3. The output feature shape is [N, 2048], N being the batch size.

2.1. I use the standard protocol of Market1501.

2.2. Choice of validation set is at your will, it's just to (1) make sure the training does not go wrong and (2) to select some super parameters. Generally, validation set should be images or persons that does not appear in the training set.

3. The difference between [0.486, 0.459, 0.408] and [0.485, 0.456, 0.406] is trivial, maybe because of the difference between 1/255 and 1/256, which I can't remember quite well now.

4. You can look into your training log, e.g. the proportion of triplets that satisfies margin > 0 and margin > your_margin (e.g. 0.3) during training. It may be due to the failure of training for some training settings, or potentially due to some mistakes.

[Zhang-Y-J]

Hi! Thanks for your reply!
We managed to find out that the low result was due to our wrong labelling. Currently we obtain 84% top 1 and 65% map for Market1501 single-query stride 1 follow the standard market1501 protocol (exclude same camera id), without re-ranking. We were trying to replicate your model however our performance is much lower than yours. Apart from the fact that our gpu only support 32 * 3 within one batch. the models should be the same. I am wondering why my model cannot get similar results to your model. Is it possible for you to provide some enlightenments to us, such as small things you did during training/test? Any suggestions would be good! Thank you very much!

[huanghoujing]

Hi, I did not tune other things not mentioned in the code. Besides keeping your setting the same as these several points, you also need to train on trainval set to obtain the reported scores.