daizuozhuo / batch-dropblock-network Goto Github PK

View Code? Open in Web Editor NEW

326.0 10.0 90.0 37 KB

Official source code of "Batch DropBlock Network for Person Re-identification and Beyond" (ICCV 2019)

License: MIT License

Python 100.00%

image-retrieval person-reidentification metric-learning pytorch

batch-dropblock-network's Introduction

Batch DropBlock Network for Person Re-identification and Beyond

Official source code of paper https://arxiv.org/abs/1811.07130

Update on 2019.3.15

Update CUHK03 results.

Update on 2019.1.29

Traning scripts are released. The best Markt1501 result is 95.3%! Please look at the training section of README.md.

Update on 2019.1.23

In-Shop Clothes Retrieval dataset and pretrained model are released!. The rank-1 result is 89.5 which is a litter bit higher than paper reported.

This paper is accepted by ICCV 2019. Please cite if you use this code in your research.

@article{dai2018batch,
  title={Batch DropBlock Network for Person Re-identification and Beyond},
  author={Dai, Zuozhuo and Chen, Mingqiang and Gu, Xiaodong and Zhu, Siyu and Tan, Ping},
  journal={arXiv preprint arXiv:1811.07130},
  year={2018}
}

Setup running environment

This project requires python3, cython, torch, torchvision, scikit-learn, tensorboardX, fire. The baseline source code is borrowed from https://github.com/L1aoXingyu/reid_baseline.

Prepare dataset

Create a directory to store reid datasets under this repo via
```bash
cd reid
mkdir data
```

For market1501 dataset, 
1. Download Market1501 dataset to `data/` from http://www.liangzheng.org/Project/project_reid.html
2. Extract dataset and rename to `market1501`. The data structure would like:
```
market1501/
    bounding_box_test/
    bounding_box_train/
    query/
```

For CUHK03 dataset,
1. Download CUHK03-NP dataset from https://github.com/zhunzhong07/person-re-ranking/tree/master/CUHK03-NP 
2. Extract dataset and rename folers inside it to cuhk-detect and cuhk-label.
For DukeMTMC-reID dataset,
Dowload from https://github.com/layumi/DukeMTMC-reID_evaluation

For In-Shop Clothes dataset,
1. Downlaod clothes dataset from http://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/bfe_models/clothes.tar
2. Extract dataset and put it to `data/` folder.

Results

Dataset	CUHK03-Label	CUHK03-Detect	DukeMTMC re-ID	Market1501	In-Shop Clothes
Rank-1	79.4	76.4	88.9	95.3	89.5
mAP	76.7	73.5	75.9	86.2	72.3
model	aliyun	aliyun]	aliyun	aliyun	aliyun

You can download the pre-trained models from the above table and evaluate on person re-ID datasets. For example, to evaluate CUHK03-Label dataset, you can download the model to './pytorch-ckpt/cuhk_label_bfe' directory and run the following commands.

Evaluate Market1501

python3 main_reid.py train --save_dir='./pytorch-ckpt/market_bfe' --model_name=bfe --train_batch=32 --test_batch=32 --dataset=market1501 --pretrained_model='./pytorch-ckpt/market_bfe/944.pth.tar' --evaluate

Evaluate CUHK03-Label

python3 main_reid.py train --save_dir='./pytorch-ckpt/cuhk_label_bfe' --model_name=bfe --train_batch=32 --test_batch=32 --dataset=cuhk-label  --pretrained_model='./pytorch-ckpt/cuhk_label_bfe/750.pth.tar' --evaluate

Evaluate In-Shop clothes

python main_reid.py train --save_dir='./pytorch-ckpt/clothes_bfe' --model_name=bfe --pretrained_model='./pytorch-ckpt/clothes_bfe/clothes_895.pth.tar' --test_batch=32 --dataset=clothes --evaluate

Training

Traning Market1501

python main_reid.py train --save_dir='./pytorch-ckpt/market-bfe' --max_epoch=400 --eval_step=30 --dataset=market1501 --test_batch=128 --train_batch=128 --optim=adam --adjust_lr

This traning command is tested on 4 GTX1080 gpus. Here is training log. You shoud get a result around 95%.

batch-dropblock-network's People

Contributors

Stargazers

Watchers

Forkers

yaoq hxl1990 whitexiezx daijucug kailigo beichen2012 zengmingm zhiqijiang wangvenn xjdupeng fukai001 irvingshu yichengwang125 theoseo zp1018 houcang hfutzzw feesics mmchen88 boosting lu-k-brant zxr8192 coallaoh dreadlord1984 137996047 millx2021 column6942 ahuwangsheng supermousse tttina jamence tiancity-nju tinyloop lzcai lijian10086 zhukkang meitianjinbu diangang yuanwei0908 faizwhb gsx0 we0091234 hsdong2012 wolfworld6 szkszk95 peterzs sft110 wangxing001 victor-gui nicolas8848 daiwc jamencetom peterzhousz eustcpl hengqujushi browserliu linar23 smallflyfly wuyangfeng igi123 yuzhijun2 pikapikazhou brianlan andrewhuman hjffily crazydemo gokulsg glingli lhc11171552 ailib decajcd lzh990711 vi1nce shengyupei kojingying guotong1223 hehaoming eduardoandrade thomaslin1990 mymuli yjhuasheng tragedyn frensher likelion-hyeonjun abandonsea emersonzc iq-scm luckyrz0

batch-dropblock-network's Issues

ValueError: expected 2D or 3D input (got 1D input)

hi~，thank you for sharing your job. when I run main_reid.py for evaluating pre-trained models on market1501 datasets, I meet a problem of ValueError: expected 2D or 3D input (got 1D input), line 68, in evaluate, feature0 = self._forward(inputs)

could you provide the scripts of training to reproduce your results? Simply run the default parameters get inferior results. Thanks

您好，我想问一下您以market1501为例，测试的指令是什么

about triplet loss

Thanks for opening your code .
I meet some issues when I train my own data by triplet loss .

shape dist_mat : 64*64
N:64
shape '[N, -1]' is invalid for input of size dist_mat[is_pos]
how to solve it ?

a question about code

hi, thank you for release your code. two lines below has different order of w_ratio and h_ratio in your code.

self.batch_crop = BatchDrop(width_ratio, height_ratio)

def init(self, h_ratio, w_ratio):

should change its order or not?

About re_ranking

It seems that it is no use to set the parameter --re_ranking = True if I want to use re-ranking, because the following code in Line 171 in main_reid.py
rank1 = reid_evaluator.evaluate(queryloader, galleryloader, queryFliploader, galleryFliploader)
doesn't pass the parameter opt.re_ranking.

bug in the code to calculate

@daizuozhuo There is something wrong in the function eval_func_gpu, the default max_rank = 50. When the numbers of images of the same id is larger than 50, it will return a wrong mAP.

test的逻辑疑问

#part branch
x = self.res_part2(x)
x = self.batch_crop(x)
triplet_feature = self.part_maxpool(x).squeeze()

楼主，你好，上述代码是您part分支进行batch drop的操作，我想问下在测试过程中为什么也要进行此操作？而当我在测试代码中去掉此操作时，他的结果和使用是完全一样的，请您指导一下，期待您的回复！！！！

The rank-1 is always 93.9%

So sorry to bother you!
The rank-1 is always 93.9% when I train the model on market-1501 . Comparing the log you have opened, there is nothing different but the configuration parameter ‘mgn_model' and 'mgn_part'. In my log, they are absent. Are they necessary? Could you give me advise and I will be appreciated. thx for your great job!

Hyperparameter config for CUB 200

Hi @daizuozhuo
Thanks a lot for sharing your amazing work.

I wanted to try to train for CUB 200. Is there a specific hyperparameter config to use for the same? or would the default in config.py?

Looking forward to your response?

Validation dataset are need ?

Hello,
I successfully executed this program, but I have some questions about the allocation of datasets.
Under normal circumstances, the training dataset and the validation dataset are carried out together.
And in your program it seems that the test dataset is used for training and also for testing.

Recall@K Code

Hi,
I think you only have CMC curve Rank-1~Rank-20 in your code,
but I don’t see the Recall@K(%) code for CUB200-2011, CARS196, Clothes, Stanford online products dataset.
Is it convenient for you to provide it?
Thank you for your reply.

请问您论文中的类激活图是怎样获得的呢？能分享下代码吗

eval_func_gpu

when using eval_func_gpu,list named results is always none

about train/test result

Thank you very much for sharing the code. I have been training less than 95.3% when I use your script to do training. I can only train about 94.5%. The environment I use is pytorch1.0.1, 4 v100, can you help me analysis? Thanks!

accuracy

Hi, I was able to execute the code correctly, but I don't know why I can't achieve the accuracy stated in the article. For the duke dataset I got a maximum accuracy of 87.2. Can I ask the reason?

"mgn_model" and " mgn_part"

There are "mgn_model" and " mgn_part" in your training log, but i can't find them in the config.py.

with provided market_953.pth.tar, unable to get 95.3 rank1 results

Hi,
Thanks a lot for your provided codes and models.

But with your provided with provided model “market_953.pth.tar”, and your code main_reid.py, I cannot get the result of rank1 95.3%.
Is there any tricks to play with inference? Thanks in advance!

I run the following command:
python3 main_reid.py train --save_dir='./pytorch-ckpt/market_bfe' --model_name=bfe --train_batch=32 --test_batch=32 --dataset=market1501 --pretrained_model='./pytorch-ckpt/market_bfe/market_953.pth.tar' --evaluate

And I get the market1501 dataset from the below link:
https://jingdongwang2017.github.io/Projects/ReID/Datasets/Market-1501.html

resuls:
=> Market1501 loaded
Dataset statistics:

subset | # ids | # images

train | 751 | 12936
query | 750 | 3368
gallery | 751 | 15913

total | 1501 | 32217

initializing model ...
load pretrained model ./pytorch-ckpt/market_bfe/market_953.pth.tar
model size: 32.27600M
Extracted features for query set: 3368 x 1536
Extracted features for gallery set: 15913 x 1536
Computing distance matrix
Computing CMC and mAP
Results ----------
mAP: 85.5%
CMC curve
Rank-1 : 94.6%
Rank-2 : 96.5%
Rank-4 : 97.4%
Rank-5 : 97.8%
Rank-8 : 98.4%
Rank-10 : 98.8%
Rank-16 : 99.3%
Rank-20 : 99.3%

hardwares and softwares:
GPU : 2080TI CUDA:10.0 pytorch:1.0.0

(2)heigth_ratio=0.0(no use of batch crop), map=85.6, rank1=94.7

does my exps right?

data_manager

您好，您能把data_manager 中关于在cuhk03数据集处理部分代码公布一下吗？谢谢！