person_reid.pytorch

Baseline model[resnet50] rank-1 achieves 95.9 on market1501!!!

Provide tool set for 首届“全国人工智能大赛”（行人重识别 Person ReID 赛项）, details related to naic2019!

Better performance, parallel training, easier usage!

This projects is my own code base for person re-identification research and competitions. The codes are heavily borrowed from 1. Based on this, I have eliminated the support of video re-id features, and added many useful training skills or structure strategy. Thanks a lot for the authors whose codes I have used.

Only based on global feature, currently!

Local features are coming, stay tuned!Welcome PR!

Better performace than deep-person-id, while more features than strong-baseline. Here are the newly features...

Features

• Lots of backbones supported(resnet, resnet_ibn, senet, densenet, mlfn, mudeep, osnet)
• Lots of losses supported(center loss, triplet loss, focal loss, ranked list loss)
• Lots of heads supported(arcface, cosface, sphere face)
• Lots of midnecks supported(single_bnneck, multi_bnneck)
• Training different parts of the model with different solvers and schedulers(for example, backbone, head or loss(center) would have different learning rates)
• Multi-gpu training(fast and robust...)

Usage

Easy to train and evaluation, as follows:

Train

• Train with single gpu

CUDA_VISIBLE_DEVICES=0 python run.py --config-file <your config file>

• Train with multi gpu

python run.py --config-file <your config file>

You can use option --name to specify the training work directory, under work_dirs/<name>, otherwise name would be choosen the default name according to your training configuration.

• Visulizing loss and learning rate

# install tensorboard first: pip install tensorboard
tensorboard --logdir <log directory> --port <port> --host <0.0.0.0>

Try to visit <server ip>:<port>

Adjust your super-parameters according to training details!

Test

• Test with groundtruth-available dataset

python run.py --config-file <your config file> --test True test.rerank False model.load_weights <checkpoint path>

If you want to test with re-rank, just set the flag test.rerank to True.

• Test with groundtruth-un-available dataset

For example, you want to get the most probably same identity images, the distance matrix will be saved as save_for_test.pkl, you would get the image names by analyzing this matrix. More details refere to tools/predict_naic2019.py.

• Visulize the test results

python run.py --config-file <your config file> --test True test.rerank False model.load_weights <checkpoint path> test.visrank True

visrank is the flag to save the visrank results to images, you can find them in the work_dirs/<name>.

MODEL_ZOO

Lots of fancy results are listed here~MODEL_ZOO

I have no time to test all of these features, any excellent config and results PR would be welcomed!

...

Any advice and discussion would be helpful for me ~!