layumi / Person_reID_baseline_pytorch

:bouncing_ball_person: Pytorch ReID: A tiny, friendly, strong pytorch implement of person re-id / vehicle re-id baseline. Tutorial 👉https://github.com/layumi/Person_reID_baseline_pytorch/tree/master/tutorial
https://www.zdzheng.xyz
MIT License
4.07k stars 1k forks source link

Any suggestions to make it work for DukeMTMC dataset? #155

Open anilsh opened 5 years ago

anilsh commented 5 years ago

Also, how can we use the same model to compute the test performance of other datasets?

Is it possible to use the provided model on other datasets? Or any suggestion about how it can be done?

layumi commented 5 years ago

Hi @anilsh ,

Yes. You could use the provided model to other datasets, since we extract the feature in the middle level (the description of the person) rather than the specific class prediction (who is the person).

However, I do not suggest to use the model trained on dataset A to dataset B directly without any adaption. Generally, it would be better to fine-tune the model on the target dataset B training data.

anilsh commented 5 years ago

Okay, thanks for the suggestion.

I am getting very low rank-1 accuracy when I try it on Duke dataset. Do you provide a trained model for duke dataset?

layumi commented 5 years ago

Sorry @anilsh Due to the demand of Duke University, I could not make any data (including trained model) public.

BabuTango commented 5 years ago

@layumi

What can be the metrics that can be fine-tuned. What are the changes we should do, to train the model on custom dataset

bmiftah commented 4 years ago

HI @layumi , In the prepare.y module , multi_query is not considered for duke , you pointed out this in the commnet as

.. for dukemtmc-reid, we do not need multi-query

why was this ? .. some multi_query ranks are reported for duke in some papers .. I see gt_box folder for market-1501 but this folder is missing for duke and hence i could't run your prepare.py to prepare data multi_query ..I appreciate if you can respond to this at your convenience

layumi commented 4 years ago

@bmiftah