zhunzhong07 / camstyle

Camera Style Adaptation for Person Re-identification CVPR 2018

Camera Style Adaptation for Person Re-identification

================================================================

Code for Camera Style Adaptation for Person Re-identification (CVPR 2018).

Preparation

Requirements: Python=3.6 and Pytorch>=0.3.0

1. Install Pytorch

2. Download dataset

   • Market-1501 [BaiduYun] [GoogleDriver]

   • DukeMTMC-reID [BaiduYun] (password: chu1) [GoogleDriver]

   • Move them to 'CamStyle/data/market (or duke)'

3. Download CamStyle Images

   • Market-1501-Camstyle [GoogleDriver] [BaiduYun] [Password:tfa2]

   • DukeMTMC-reID-CamStyle [GoogleDriver] [BaiduYun] [Password:kibf]

   • Move them to 'CamStyle/data/market (or duke)/bounding_box_train_camstyle'

CamStyle Generation

You can generate CamStyle imgaes with CycleGAN-for-CamStyle

Training and test re-ID model

1. IDE

# For Market-1501
python main.py -d market --logs-dir logs/market-ide
# For Duke
python main.py -d duke --logs-dir logs/duke-ide

2. IDE + CamStyle

# For Market-1501
python main.py -d market --logs-dir logs/market-ide-camstyle --camstyle 46
# For Duke
python main.py -d duke --logs-dir logs/duke-ide--camstyle --camstyle 46

3. IDE + CamStyle + Random Erasing[4]

# For Market-1501
python main.py -d market --logs-dir logs/market-ide-camstyle-re --camstyle 46 --re 0.5
# For Duke
python main.py -d duke --logs-dir logs/duke-ide--camstyle-re --camstyle 46 --re 0.5

4. IDE + CamStyle + Random Erasing[4] + re-ranking[3]

# For Market-1501
python main.py -d market --logs-dir logs/market-ide-camstyle-re --camstyle 46 --re 0.5 --rerank
# For Duke
python main.py -d duke --logs-dir logs/duke-ide--camstyle-re --camstyle 46 --re 0.5 --rerank

 

Results

	Market-1501		Duke
Methods	Rank-1	mAP	Rank-1	mAP
IDE	85.6	65.8	72.3	51.8
IDE+CamStyle	88.1	68.7	75.2	53.4
IDE+CamStyle+Random Erasing	89.4	71.5	78.3	57.6

References

• [1] Our code is conducted based on open-reid

• [2] Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, ICCV 2017

• [3] Re-ranking Person Re-identification with k-reciprocal Encoding. CVPR 2017.

• [4] Random Erasing Data Augmentation. Arxiv 2017.

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{zhong2018camera,
title={Camera Style Adaptation for Person Re-identification},
author={Zhong, Zhun and Zheng, Liang and Zheng, Zhedong and Li, Shaozi and Yang, Yi},
booktitle={CVPR},
year={2018}
}

Contact me

If you have any questions about this code, please do not hesitate to contact me.

Zhun Zhong

camstyle's Issues

about the test

The data augmentation is used for training. When testing, do you use the generated images? Or the test is just like the normal test?

ide without camstyle got higher mAP(my batch-size is 64)

Sorry to interrupt again, when I ran "python main.py -d market --logs-dir logs/market-ide" I got this

so I just comment the camstyle part in main.py(sepecficly, the function named get_data and the function named main), then I got the result like this
mAP : 70.2%
top-1 : 87.9%
then I got everything back, I ran "python main.py -d market --logs-dir logs/market-ide-camstyle --camstyle 22", I suppose the number behind --camstyle divided by batch-size is the ratio between fake and real images.
However I got the mAP of 68.6% and top-1 of 87.5%, so the result was deteriorating after I added the camstyle image. I hope you can solve my problem. Thank you.

Baseline (IDE) accuracy 1% off

Hello,
Thanks for sharing the code!
I ran your code without any modification for re-id baseline without using camstyle.
I was only able to get 84.3% instead of 85.6% in Readme.

Is there any other hyperparameter setting that I have to change to achieve 85% accuracy for baseline IDE?
Any idea where that 1 % gap comes from?

Thank you!

RuntimeError: CUDA error: out of memory

Hi, when I run the comand
“python main.py -d market --logs-dir logs/market-ide-camstyle --camstyle 46”

I get errot
“RuntimeError: CUDA error: out of memory”
Can you help me with your answer?

当运行reid tripletloss时主函数应当怎么修改

我修改了loss选择与选择了输出为特征，但是特有的数据加载方式应该如何修改，我在reid/utils/data中找到了sampler.py与prerocessor.py文件 但我不知如何在主函数中使用

something wrong with my computergot like batch_size=0

(pytorch) D:\test03>python main.py -d market --logs-dir logs/market-ide
Market dataset loaded
subset | # ids | # images

train | 0 | 0
query | 0 | 0
gallery | 0 | 0
camstyle | 0 | 0
Traceback (most recent call last):
File "main.py", line 198, in
main(parser.parse_args())
File "main.py", line 83, in main
args.width, args.batch_size, args.camstyle, args.re, args.workers)
File "main.py", line 69, in get_data
shuffle=True, pin_memory=True, drop_last=True)
File "C:\ProgramData\Anaconda3\envs\pytorch\lib\site-packages\torch\utils\data\dataloader.py", line 437, in init
batch_sampler = BatchSampler(sampler, batch_size, drop_last)
File "C:\ProgramData\Anaconda3\envs\pytorch\lib\site-packages\torch\utils\data\sampler.py", line 128, in init
"but got batch_size={}".format(batch_size))
ValueError: batch_size should be a positive integeral value, but got batch_size=0

when I ran 'python main.py -d market --logs-dir logs/market-ide' on my computer, I got the information above,what can I do to solove this problem? thank you.

ValueError: num_samples should be a positive integer value, but got num_samples=0

root@833c2b940c92:/home/CamStyle# CUDA_VISIBLE_DEVICES=0 python main.py -d market --logs-dir logs/market-ide-camstyle --camstyle 46
Market dataset loaded
subset | # ids | # images

train | 751 | 12936
query | 750 | 3368
gallery | 751 | 15913
camstyle | 0 | 0
Traceback (most recent call last):
File "main.py", line 201, in
main(parser.parse_args())
File "main.py", line 86, in main
args.width, args.batch_size, args.camstyle, args.re, args.workers)
File "main.py", line 72, in get_data
shuffle=True, pin_memory=True, drop_last=True)
File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 176, in init
sampler = RandomSampler(dataset)
File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/sampler.py", line 66, in init
"value, but got num_samples={}".format(self.num_samples))
ValueError: num_samples should be a positive integer value, but got num_samples=0

I want to run your code on translated market-1501 image +original images. Is there any way to run the code?

hello,I'm confused with your train strategy

In the paper,you viewed the real and fake images equally,as to the loss function, when defining loss function , why you considered it into two parts ?

could you upload pretrained model

Hi, would you please release the model you trained for these works?

Thanks a lot.

not class methods!

when I run the comand

python main.py -d duke --logs-dir logs/duke-ide

I get errot

  File "main.py", line 42, in get_data
    T.Resize((height, width), interpolation=3),
AttributeError: module 'reid.utils.data.transforms' has no attribute 'Resize'

So I saw you reid.utils.data.transforms module，in face ,not Resize attribute.

RuntimeError: Invalid index in scatter at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:549

Hi. The total number of categories in the training set is 751, but the category ID read from the training set is beyond this range. How to solve this problem? Thanks.

百度网盘中的DukeMTMC-reID数据集已经无法下载了

result question

Why is the Rank-1 and mAP of the IDE+CamStyle experiment when I directly use the code and the author's picture lower than the Rank-1 and mAP of the IDE experiment?
In IDE + CamStyle, I just run "python main.py -d duke --logs-dir logs/duke-ide--camstyle --camstyle 46" with "DukeMTMC-reID-CamStyle".
In IDE, I just run "python main.py -d duke --logs-dir logs/duke-ide".
What more do I need to do?

No module called dominate

"html.py", line 1, in
import dominate
ImportError: No module named dominate

i can't find dominate file in your project, can you help me with this?

a error on my computer

aaron@PC:~/reid/CamStyle$ python main.py -d market --logs-dir logs/market-ide
Market dataset loaded
subset | # ids | # images

train | 0 | 0
query | 0 | 0
gallery | 0 | 0
camstyle | 0 | 0
Traceback (most recent call last):
File "main.py", line 198, in
main(parser.parse_args())
File "main.py", line 138, in main
trainer.train(epoch, train_loader, optimizer)
File "/home/aaron/reid/CamStyle/reid/trainers.py", line 28, in train
for i, inputs in enumerate(data_loader):
File "/home/aaron/anaconda3/lib/python3.5/site-packages/torch/utils/data/dataloader.py", line 417, in iter
return DataLoaderIter(self)
File "/home/aaron/anaconda3/lib/python3.5/site-packages/torch/utils/data/dataloader.py", line 242, in init
self._put_indices()
File "/home/aaron/anaconda3/lib/python3.5/site-packages/torch/utils/data/dataloader.py", line 290, in _put_indices
indices = next(self.sample_iter, None)
File "/home/aaron/anaconda3/lib/python3.5/site-packages/torch/utils/data/sampler.py", line 119, in iter
for idx in self.sampler:
File "/home/aaron/anaconda3/lib/python3.5/site-packages/torch/utils/data/sampler.py", line 50, in iter
return iter(torch.randperm(len(self.data_source)).long())
RuntimeError: invalid argument 1: must be strictly positive at /opt/conda/conda-bld/pytorch_1523242347739/work/torch/lib/TH/generic/THTensorMath.c:2247
......
when I ran 'python main.py -d market --logs-dir logs/market-ide' on my computer, I got the information above, so I checked the code. I found in the line 36 of evaluators.py, you used 'for i, (imgs, fnames, pids, _) in enumerate(data_loader):' to process the data_loader, while in line 28 of trainers.py you used 'for i, inputs in enumerate(data_loader):', then you defined a function named _parse_data to process this 'inputs', then I got the error. So I was wondering if you have the same problem, and am I supposed to change the code there?
(I tried on pytorch 0.3.1 on both python 3.5 and python 3.6, and also I thought maybe it was my pytorch's problem?

batch size and epoch

Hi, How much should epoch be set when batch size is set to 3. Thanks.

hi,can u share me the batchsize and epoch num when u train generate model? thanks

When i use train_market.sh to train generate model,i don't kown how to set best batchsize and epoch number. can u give me your training parameter（batchsize and opech number）？
With the default parameter ，the quality of generated image is not very well,so i think it might be parameter problem.

请问您的GAN模型是用什么数据训练的呢？也是用market和duke吗？

Identity loss weight in CircleGAN?

What's the times of identity loss weight smaller than reconstruction loss?