Baseline Code (with bottleneck) for Person-reID (pytorch). It is consistent with the new baseline result in Beyond Part Models: Person Retrieval with Refined Part Pooling and Camera Style Adaptation for Person Re-identification.
We arrived Rank@1=88.24%, mAP=70.68% only with softmax loss.
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: 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: 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 16GB Memory. You may try apply smaller batchsize and choose the smaller learning rate (for stability) to run.
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 easily arrive Rank@1=89.91% mAP=73.58%. Trained DenseNet-121 model can be found at GoogleDrive.(Note that ResNet-50 is a more common choice as the baseline.)
What's new： Trained ResNet-50 model is available at GoogleDrive.
What's new: Re-ranking is added to evaluation. The re-ranked result is 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.
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+
(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
Because pytorch and torchvision are ongoing projects.
Here we noted that our code is tested based on Pytorch 0.3.0/0.4.0 and Torchvision 0.2.0.
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.
To save trained model, we make a dir.
mkdir model 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 --which_epoch 59--gpu_ids which gpu to run.
--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.
The model is based on Resnet50. Input images are resized to 256x128. Here we just show some results.
Note that the result may contain around 1% bias.(For example, 50th-epoch model can be better.)
| BatchSize | Dropout | Rank@1 | mAP | Note |
|---|---|---|---|---|
| 16 | 0.5 | 86.67 | 68.19 | |
| 32 | 0.5 | 87.98 | 69.38 | |
| 32 | 0.5 | 88.24 | 70.68 | test with 288x144 |
| 32 | 0.5 | 89.13 | 73.50 | train with random erasing and test with 288x144 |
| 32 | 0.5 | 87.14 | 68.90 | 0.1 color jitter |
| 64 | 0.5 | 86.82 | 67.48 | |
| 64 | 0.5 | 85.78 | 65.97 | 0.1 color jitter |
| 64 | 0.5 | 85.42 | 65.29 | 0.4 color jitter |
| 64 | 0.75 | 84.86 | 66.06 | |
| 96 | 0.5 | 86.05 | 67.03 | |
| 96 | 0.75 | 85.66 | 66.44 |
Test with 144x288, dropout rate is 0.5
| BatchSize | Bottleneck | Rank@1 | mAP | Note |
|---|---|---|---|---|
| 32 | 256 | 87.26 | 69.92 | |
| 32 | 512 | 88.24 | 70.68 | |
| 32 | 1024 | 84.29 | 64.00 |
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}
}