A tiny, friendly, strong baseline code for Person-reID (based on pytorch).
-
Strong. It is consistent with the new baseline result in several top-conference works, e.g., Beyond Part Models: Person Retrieval with Refined Part Pooling(ECCV18) and Camera Style Adaptation for Person Re-identification(CVPR18). We arrived Rank@1=88.24%, mAP=70.68% only with softmax loss.
-
Small. With fp16, our baseline could be trained with only 2GB GPU memory.
-
Friendly. You may use the off-the-shelf options to apply many state-of-the-art tricks in one line. Besides, if you are new to person re-ID, you may check out our Tutorial first (8 min read) đź‘Ť .
Now we have supported:
- Float16 to save GPU memory based on apex
- Part-based Convolutional Baseline(PCB)
- Multiple Query Evaluation
- Re-Ranking
- Random Erasing
- ResNet/DenseNet
- Visualize Training Curves
- Visualize Ranking Result
Here we provide hyperparameters and architectures, that were used to generate the result. Some of them (i.e. learning rate) are far from optimal. Do not hesitate to change them and see the effect.
P.S. With similar structure, we arrived Rank@1=87.74% mAP=69.46% with Matconvnet. (batchsize=8, dropout=0.75) You may refer to Here. Different framework need to be tuned in a different way.
What's new: FP16 has been added. It can be used by simply added --fp16. You need to install apex and update your pytorch to 1.0.
Float16 could save about 50% GPU memory usage without accuracy drop. Our baseline could be trained with only 2GB GPU memory.
python train.py --fp16What's new: Visualizing ranking result is added.
python prepare.py
python train.py
python test.py
python demo.py --query_index 777What's new: Multiple-query Evaluation is added. The multiple-query result is about Rank@1=91.95% mAP=78.06%.
python prepare.py
python train.py
python test.py --multi
python evaluate_gpu.pyWhat's new: Â PCB is added. You may use '--PCB' to use this model. It can achieve around Rank@1=92.73% mAP=78.16%. I used a GPU (P40) with 24GB Memory. You may try apply smaller batchsize and choose the smaller learning rate (for stability) to run. (For example, --batchsize 32 --lr 0.01 --PCB)
python train.py --PCB --batchsize 64 --name PCB-64
python test.py --PCB --name PCB-64What's new: You may try evaluate_gpu.py to conduct a faster evaluation with GPU.
What's new: You may apply '--use_dense' to use DenseNet-121. It can arrive around Rank@1=89.91% mAP=73.58%.
What's new: Re-ranking is added to evaluation. The re-ranked result is about Rank@1=90.20% mAP=84.76%.
What's new: Random Erasing is added to train.
What's new: I add some code to generate training curves. The figure will be saved into the model folder when training.
I re-trained several models, and the results may be different with the original one. Just for a quick reference, you may directly use these models. The download link is Here.
| Methods | Rank@1 | mAP | Reference |
|---|---|---|---|
| [ResNet-50] | 88.84% | 71.49% | python train.py --train_all |
| [DenseNet-121] | 90.11% | 73.51% | python train.py --name ft_net_dense --use_dense --train_all |
| [PCB] | 92.64% | 77.47% | python train.py --name PCB --PCB --train_all --lr 0.02 |
| [ResNet-50 (fp16)] | 88.27% | 71.20% | python train.py --name fp16 --fp16 --train_all |
You may learn more from model.py.
We add one linear layer(bottleneck), one batchnorm layer and relu.
- Python 3.6
- GPU Memory >= 6G
- Numpy
- Pytorch 0.3+
- [Optional] apex (for float16)
(Some reports found that updating numpy can arrive the right accuracy. If you only get 50~80 Top1 Accuracy, just try it.) We have successfully run the code based on numpy 1.12.1 and 1.13.1 .
- Install Pytorch from http://pytorch.org/
- Install Torchvision from the source
git clone https://github.com/pytorch/vision
cd vision
python setup.py install
- [Optinal] You may skip it. Install apex from the source
git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext
Because pytorch and torchvision are ongoing projects.
Here we noted that our code is tested based on Pytorch 0.3.0/0.4.0/0.5.0/1.0.0 and Torchvision 0.2.0/0.2.1 .
Download Market1501 Dataset
Preparation: Put the images with the same id in one folder. You may use
python prepare.pyRemember to change the dataset path to your own path.
Futhermore, you also can test our code on DukeMTMC-reID Dataset. Our baseline code is not such high on DukeMTMC-reID Rank@1=64.23%, mAP=43.92%. Hyperparameters are need to be tuned.
Train a model by
python train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32 --data_dir your_data_path--gpu_ids which gpu to run.
--name the name of model.
--data_dir the path of the training data.
--train_all using all images to train.
--batchsize batch size.
--erasing_p random erasing probability.
Train a model with random erasing by
python train.py --gpu_ids 0 --name ft_ResNet50 --train_all --batchsize 32 --data_dir your_data_path --erasing_p 0.5Use trained model to extract feature by
python test.py --gpu_ids 0 --name ft_ResNet50 --test_dir your_data_path --batchsize 32 --which_epoch 59--gpu_ids which gpu to run.
--batchsize batch size.
--name the dir name of trained model.
--which_epoch select the i-th model.
--data_dir the path of the testing data.
python evaluate.pyIt will output Rank@1, Rank@5, Rank@10 and mAP results.
You may also try evaluate_gpu.py to conduct a faster evaluation with GPU.
For mAP calculation, you also can refer to the C++ code for Oxford Building. We use the triangle mAP calculation (consistent with the Market1501 original code).
python evaluate_rerank.pyIt may take more than 10G Memory to run. So run it on a powerful machine if possible.
It will output Rank@1, Rank@5, Rank@10 and mAP results.
As far as I know, the following papers may be the first two to use the bottleneck baseline. You may cite them in your paper.
@article{DBLP:journals/corr/SunZDW17,
author = {Yifan Sun and
Liang Zheng and
Weijian Deng and
Shengjin Wang},
title = {SVDNet for Pedestrian Retrieval},
booktitle = {ICCV},
year = {2017},
}
@article{hermans2017defense,
title={In Defense of the Triplet Loss for Person Re-Identification},
author={Hermans, Alexander and Beyer, Lucas and Leibe, Bastian},
journal={arXiv preprint arXiv:1703.07737},
year={2017}
}