search

Cysu / dgd_person_reid

Domain Guided Dropout for Person Re-identification
http://arxiv.org/abs/1604.07528
231 stars 94 forks source link

I use your model but get better results,why? #26

Closed liaowang0125 closed 7 years ago

liaowang0125 commented 7 years ago

cuhk03 top-1 93.8%

cuhk01 top-1 77.5%

prid top-1 69.0%

viper top-1 38.3%

3dpes top-1 53.5%

ilids top-1 62.4%

Cysu commented 7 years ago

We randomly split the datasets for training and testing. Our trained model may not be directly used for evaluation on these six datasets. But you can finetune it on other datasets, or just train from scratch by yourself.

liaowang0125 commented 7 years ago

But I find that your paper show the JSTL+DGD results on six datasets without finetune, i think the results i get is that situation,right? And diffrent results is caused by randomly split the datasets. @Cysu

Cysu commented 7 years ago

Right. What I mean is that you can safely use our trained model as an initial point for finetuning on other datasets, for example, market 1501. The higher results are caused by randomly splits.

liaowang0125 commented 7 years ago

Okay,i understand.Thank you very much.

nullmax commented 6 years ago

Have you modified the code? I ran the code as the readme without any modification, but get terrible results. For example, the cuhk03 keep the following results in individual experiments, JSTL, JSTL+DGD and FT-(JSTL_DGD) top-1 9.0% top-5 45.4% top-10 88.7% top-20 95.8% The other datasets keep the same bad result.