Bag of Tricks and A Strong ReID Baseline

Bag of Tricks and A Strong Baseline for Deep Person Re-identification. CVPRW2019, Oral.

A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification. IEEE Transactions on Multimedia (Accepted).

[Journal Version(TMM)] [PDF] [Slides] [Poster]

News! Based on the strong baseline, we won 3rd place on AICity Challenge 2020. [PDF] [Code]

News! Our journal version has been accepted by IEEE Transactions on Multimedia.

We are very grateful for your contribution to our project and hope that this project can help your research or work.

The codes are expanded on a ReID-baseline , which is open sourced by our co-first author Xingyu Liao.

Another re-implement is developed by python2.7 and pytorch0.4. [link]

A tiny repo with simple re-implement. [link]

Our baseline also achieves great performance on Vehicle ReID task! [link]

With Ranked List loss(CVPR2019)[link], our baseline can achieve better performance. [link]

@InProceedings{Luo_2019_CVPR_Workshops,
author = {Luo, Hao and Gu, Youzhi and Liao, Xingyu and Lai, Shenqi and Jiang, Wei},
title = {Bag of Tricks and a Strong Baseline for Deep Person Re-Identification},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}

@ARTICLE{Luo_2019_Strong_TMM, 
author={H. {Luo} and W. {Jiang} and Y. {Gu} and F. {Liu} and X. {Liao} and S. {Lai} and J. {Gu}}, 
journal={IEEE Transactions on Multimedia}, 
title={A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification}, 
year={2019}, 
pages={1-1}, 
doi={10.1109/TMM.2019.2958756}, 
ISSN={1941-0077}, 
}

Authors

• Hao Luo
• Youzhi Gu
• Xingyu Liao
• Shenqi Lai

We support

• [x] easy dataset preparation
• [x] end-to-end training and evaluation
• [x] high modular management
• [x] speed up inference [link]
• [x] support multi-gpus training [link]

Bag of tricks

• Warm up learning rate
• Random erasing augmentation
• Label smoothing
• Last stride
• BNNeck
• Center loss

TODO list

In the future, we will

• [] support more datasets
• [] support more models
• [] explore more tricks

Pipeline

Results (rank1/mAP)

model(Market1501)

model(DukeMTMC-reID)

Get Started

The designed architecture follows this guide PyTorch-Project-Template, you can check each folder's purpose by yourself.

1. cd to folder where you want to download this repo

2. Run git clone https://github.com/michuanhaohao/reid-strong-baseline.git

3. Install dependencies:

   • pytorch>=0.4
   • torchvision
   • ignite=0.1.2 (Note: V0.2.0 may result in an error)
   • yacs

4. Prepare dataset

   Create a directory to store reid datasets under this repo or outside this repo. Remember to set your path to the root of the dataset in config/defaults.py for all training and testing or set in every single config file in configs/ or set in every single command.

   You can create a directory to store reid datasets under this repo via

   cd reid-strong-baseline
   mkdir data
   

   （1）Market1501

   • Download dataset to data/ from http://www.liangzheng.org/Project/project_reid.html
   • Extract dataset and rename to market1501. The data structure would like:

   data
       market1501 # this folder contains 6 files.
           bounding_box_test/
           bounding_box_train/
           ......
   

   （2）DukeMTMC-reID

   • Download dataset to data/ from https://awesomeopensource.com/project/layumi/DukeMTMC-reID_evaluation#download-dataset
   • Extract dataset and rename to dukemtmc-reid. The data structure would like:

   data
       dukemtmc-reid
       	DukeMTMC-reID # this folder contains 8 files.
           	bounding_box_test/
           	bounding_box_train/
           	......
   

5. Prepare pretrained model if you don't have

   （1）ResNet

   from torchvision import models
   models.resnet50(pretrained=True)
   

   （2）Senet

   import torch.utils.model_zoo as model_zoo
   model_zoo.load_url('the pth you want to download (specific urls are listed in  ./modeling/backbones/senet.py)')
   

   Then it will automatically download model in ~/.torch/models/, you should set this path in config/defaults.py for all training or set in every single training config file in configs/ or set in every single command.

   （3）ResNet_IBN_a

   You can download the ImageNet pre-trained weights from here [link]

   （4）Load your self-trained model If you want to continue your train process based on your self-trained model, you can change the configuration PRETRAIN_CHOICE from 'imagenet' to 'self' and set the PRETRAIN_PATH to your self-trained model. We offer Experiment-pretrain_choice-all_tricks-tri_center-market.sh as an example.

6. If you want to know the detailed configurations and their meaning, please refer to config/defaults.py. If you want to set your own parameters, you can follow our method: create a new yml file, then set your own parameters. Add --config_file='configs/your yml file' int the commands described below, then our code will merge your configuration. automatically.

Train

You can run these commands in .sh files for training different datasets of differernt loss. You can also directly run code sh *.sh to run our demo after your custom modification.

1. Market1501, cross entropy loss + triplet loss

python3 tools/train.py --config_file='configs/softmax_triplet.yml' MODEL.DEVICE_ID "('your device id')" DATASETS.NAMES "('market1501')" OUTPUT_DIR "('your path to save checkpoints and logs')"

2. DukeMTMC-reID, cross entropy loss + triplet loss + center loss

python3 tools/train.py --config_file='configs/softmax_triplet_with_center.yml' MODEL.DEVICE_ID "('your device id')" DATASETS.NAMES "('dukemtmc')" OUTPUT_DIR "('your path to save checkpoints and logs')"

Test

You can test your model's performance directly by running these commands in .sh files after your custom modification. You can also change the configuration to determine which feature of BNNeck is used and whether the feature is normalized (equivalent to use Cosine distance or Euclidean distance) for testing.

Please replace the data path of the model and set the PRETRAIN_CHOICE as 'self' to avoid time consuming on loading ImageNet pretrained model.

1. Test with Euclidean distance using feature before BN without re-ranking,.

python3 tools/test.py --config_file='configs/softmax_triplet_with_center.yml' MODEL.DEVICE_ID "('your device id')" DATASETS.NAMES "('market1501')" TEST.NECK_FEAT "('before')" TEST.FEAT_NORM "('no')" MODEL.PRETRAIN_CHOICE "('self')" TEST.WEIGHT "('your path to trained checkpoints')"

2. Test with Cosine distance using feature after BN without re-ranking,.

python3 tools/test.py --config_file='configs/softmax_triplet_with_center.yml' MODEL.DEVICE_ID "('your device id')" DATASETS.NAMES "('market1501')" TEST.NECK_FEAT "('after')" TEST.FEAT_NORM "('yes')" MODEL.PRETRAIN_CHOICE "('self')" TEST.WEIGHT "('your path to trained checkpoints')"

3. Test with Cosine distance using feature after BN with re-ranking

python3 tools/test.py --config_file='configs/softmax_triplet_with_center.yml' MODEL.DEVICE_ID "('your device id')" DATASETS.NAMES "('dukemtmc')" TEST.NECK_FEAT "('after')" TEST.FEAT_NORM "('yes')" MODEL.PRETRAIN_CHOICE "('self')" TEST.RE_RANKING "('yes')" TEST.WEIGHT "('your path to trained checkpoints')"