search

layumi / Person-reID_GAN

ICCV2017 Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro
https://arxiv.org/abs/1701.07717
MIT License
320 stars 89 forks source link

Baseline result is very high #6

Closed FlyHighest closed 6 years ago

FlyHighest commented 6 years ago

I use train_id_net_res_market_new ， without GAN data, dataset Market1501, and the final result is r1-5: 0.8216 0.8679 0.8934 0.9068 0.9192 mAP : 0.6038. The evaluation code is '*faster', single query. Is there anyone else who has run the baseline code?

layumi commented 6 years ago

Hi @FlyHighest Do you change the baseline code? I notice that the small batchsize can improve the result.

FlyHighest commented 6 years ago

Thanks for your reply. i change batch size to 10. But the same batch size (10) fails to improve LSRO. I got rank1 0.7993.

layumi commented 6 years ago

@FlyHighest It is possible. We also notice this observation. It also appears in this ICLR paper. For some small datasets, small batchsize can do help to aviod the overfitting.

In fact, we are planning to discuss and investigate the result in our journal paper. P.S. On CUBird dataset, LSRO still works.

Thank you for your kindly attention.

FlyHighest commented 6 years ago

Thank you. It helps a lot. I'm looking forward to your new papers.