baseline results about alignedreid-re-production-pytorch

[huanghoujing]

GL-NF only uses Triplet Loss with Global distance. The difference to open-reid triplet loss example includes

• whether to use standard Market1501 train/test split, false in open-reid, but true here
• the type of data augmentation, mirror + cropping in open-reid, only mirror here
• batch size may be different, here 128 images
• whether to use an extra embedding layer (128-dim) after average pooling (2048-dim), used in open-reid, but here not
• whether to normalize the feature before computing triplet loss and for testing, done in GL-NF. This has little influence when triplet loss is only trained with global distance.
• the epoch at which the learning rate starts to decay, epoch 100 for open-reid and 75 here

Thanks :)

[marvis]

Thanks for sharing your training details!

However I have different experiment results comparing with your fifth item. After feature normalization before computing triplet loss, the CMC score decreases ~8%. And I find that with or without extra embedding layer doesn't affect the accuracy. So I doubt whether normalization caused the different baseline result?

Best,

[huanghoujing]

1. I found that for Global Distance + Triplet Loss, i.e. the commonly used triplet loss paradigm, normalizing feature or not has little influence.
2. Without normalization, I found it ok to train Global Distance + Local Distance + Triplet Loss. So later experiments only consider not normalizing features.
3. By decreases ~8%, do you mean GL-LL-NF-LHSFLD-TWLD is worse than GL-LL-NNF-LHSFLD-TWLD by ~8 CMC Rank-1 points?
4. BTW, the percentage of satisfying margin, i.e. distance(anchor, positive) + margin < distance(anchor, negative) is a good signal to check the status of convergence. It should be at least over 90% for both global and local distance. Do you find something abnormal in this signal at the end of training?

[huanghoujing]

With more experiments, I find that for vanilla triplet loss, with and without normalizing feature indeed makes difference, the later being better. Some results are in the provided Excel file.

Thank you for discussion.