Basic Person ReID Baseline and Project Template

A Basic Person ReID Baseline and a PyTorch Template for NTU ROSE Person ReID Project.

I am not a big fan of pytorch ignite(Too high level). So I have rewrite L1aoXingyu's reid_baseline following the basic pytorch training and testing logtic flow. As a basic reid baseline, I remove most of tricks and custom-made scheduler, except the bash hard triplet loss and random erasing. Evething elso are all pytorch native build-in functions.

Requirements

• python 3
• pytorch 1.0 + torchvision
• yacs Yet Another Configuration System
• fire Automatically generating command line interfaces (CLIs)

Install all dependences libraries

pip3 install -r requirements.txt

Configs

Use different yaml config files for different experiment settings. All the config files are store in folder config. Please use different OUTPUT_DIR names for different experiments to avoid conflit and accidentally files overwritten.

Datasets

This code support CUHK03, Market1501, DukeMTMC and MSMT17 datasets. All these dataset should be defined in the DATASETS.NAMES of the config file, our code will be download the corresponding dataset automatically (into the datasets folder). As this fuction require access to Google Drive, it will not work in China. Currently support:

• CUHK03
• Market1501
• DukeMTMC
• MSMT17

Training:

• Batch Size 128 uses around 12.01G GPU memory (Only recommend for Titian GPU and above on server interface)
• Batch Size 64 is the most suitable size for GTX 1080ti

python train.py ./config/market_softmax.yaml

### Change GPU
python train.py ./config/market_softmax.yaml --DEVICE=cuda:5

Testing:

• Default testing batch size is 256. Reduce to accommodate your GPU memory size.

### No Re-Ranking
python test.py ./config/market_softmax.yaml

### Change GPU
python test.py ./config/market_softmax.yaml --DEVICE=cuda:5

### With Re-Ranking
python test.py ./config/market_softmax.yaml --RE_RANKING=True

Testing Cross Dataset:

### Market1501 -> DukeMTMC
python test_cross_dataset.py ./config/market_softmax.yaml DukeMTMC

Results

Batch Size 128: Rank1 (mAP)

	Softmax	Softmax+Triplet	Softmax+Re-ranking	Softmax+Triplet+Re-ranking
CUHK03	61.8 (58.7)	63.6 (60.2)	68.2 (70.0)	72.6 (73.9)
Market1501	91.3 (77.8)	92.8 (82.0)	90.6 (85.7)	93.3 (90.1)
DukeMTMC	84.1 (67.7)	86.2 (73.0)	85.3 (79.6)	88.2 (83.5)
MSMT17	71.6 (43.9)	74.0 (47.5)	-	-

Batch Size 64: Rank1 (mAP)

	Softmax	Softmax+Triplet	Softmax+Re-ranking	Softmax+Triplet+Re-ranking
CUHK03	56.1 (52.4)	65.6 (61.8)	64.2 (64.9)	74.6 (75.5)
Market1501	91.6 (78.7)	93.2 (82.0)	90.8 (85.9)	93.8 (90.2)
DukeMTMC	83.4 (66.6)	86.4 (72.4)	84.9 (79.5)	88.1 (83.0)
MSMT17	69.0 (40.1)	73.9 (46.4)	-	-

File and Folder Structure

├──  checkpoint - here's store all the training models checkpoints and testing results
│    └── Market1501
│        └── Softmax_BS64
│            └── log.txt                 - training log
│            └── ResNet50_epo120.pth     - saved model checkpoint parameters
│            └── result.txt              - testing result
│            └── result_re-ranking.txt   - testing result with re-ranking
│ 
│
├──  config
│    └── defaults.py  - here's the default config file.
│    └── market_softmax.yml  - here's the specific config file for specific model or dataset.
│ 
│
├──  data_loader  
│    └── datasets_importer  - here's the datasets folder that is responsible for all data handling.
│        └── BaseDataset.py  - Generate and show basic statistics of the dataset in terminal.
│        └── ImageDataset.py  - PIL read the images and gernerate PyTorch Dataset Object
│        └── market1501.py  - Data handler for dataset Market1501
│
│    └── transforms  - here's the data preprocess folder is responsible for all data augmentation.
│        └── transforms.py  - initialization of data transformation of the network
│        └── RandomErasing.py  - Custom-made RandomErasing process for data augmentation
│ 
│    └── samplers  - here's the id samplering function for triplet training
│        └── triplet_sampler.py
│ 
│    └── data_loader.py  - here's the file to make dataloader.
│
│
├──  datasets  
│    └── PersonReID_Dataset_Downloader.py  - here's the file to automatic download the dataset
│    └── Market1501  - here will be the folder storing the downloaded dataset
│
│
├──  evaluation
│   ├── evaluation.py   - this file to compute the CMC and mAP result.
│   └── re_ranking.py   - this file is the re_ranking function.
│
│
├── logger  - this folder is to create a logger and store the training process.
│
│
├── loss  - this folder is the loss function for the network.
│   └── make_loss.py
│   └── triplet_loss.py  - Custom-made Triplet Loss function
│  
│
├── models  - this folder contains models of the project.
│   └── BasicModule.py     - Re-package the Pytorch Model with save and load models function
│   └── ResNet50.py        - Model with ResNet50 as backbone
│
│
├── optimizer - this folder contains optimizer of the project.
│
├── scheduler - this folder contains learning rate scheduler
|
├── utils       
│   └── check_jupyter_run.py - if it is running on the jupyter use the notebook version of tqdm
│   
│ 
├── train.py                - here's the train the network
│    
└── test.py                 - here's the test the network performance   
│
└── test_cross_dataset.py	- test the performance in cross-dataset scenario

Issues

• Re-Ranking is not working on MSMT17. Currently under investigating