Self-paced Contrastive Learning (SpCL)

The official repository for Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID, which is accepted by NeurIPS-2020.

SpCL

Updates

[2020-10-13] All trained models for the camera-ready version have been updated, see Trained Models for details.

[2020-09-25]

SpCL

[2020-07-01] We did the code refactoring to support distributed training, stronger performances and more features. Please see OpenUnReID.

Requirements

Installation

git clone https://github.com/yxgeee/SpCL.git
cd SpCL
python setup.py develop

Prepare Datasets

cd examples && mkdir data

Download the person datasets Market-1501, MSMT17, PersonX, and the vehicle datasets VehicleID, VeRi-776, VehicleX. Then unzip them under the directory like

SpCL/examples/data
├── market1501
│   └── Market-1501-v15.09.15
├── msmt17
│   └── MSMT17_V1
├── personx
│   └── PersonX
├── vehicleid
│   └── VehicleID -> VehicleID_V1.0
├── vehiclex
│   └── AIC20_ReID_Simulation -> AIC20_track2/AIC20_ReID_Simulation
└── veri
    └── VeRi -> VeRi_with_plate

Prepare ImageNet Pre-trained Models for IBN-Net

When training with the backbone of IBN-ResNet, you need to download the ImageNet-pretrained model from this link and save it under the path of

logs/pretrained/

shell
mkdir logs && cd logs
mkdir pretrained

SpCL/logs
└── pretrained
    └── resnet50_ibn_a.pth.tar

Training

We utilize 4 GTX-1080TI GPUs for training. Note that

• The training for

  SpCL

  is end-to-end, which means that no source-domain pre-training is required.
• use

  --iters 400

  (default) for Market-1501 and PersonX datasets, and

  --iters 800

  for MSMT17, VeRi-776, VehicleID and VehicleX datasets;
• use

  --width 128 --height 256

  (default) for person datasets, and

  --height 224 --width 224

  for vehicle datasets;
• use

  -a resnet50

  (default) for the backbone of ResNet-50, and

  -a resnet_ibn50a

  for the backbone of IBN-ResNet.

Unsupervised Domain Adaptation

To train the model(s) in the paper, run this command:

shell
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python examples/spcl_train_uda.py \
  -ds $SOURCE_DATASET -dt $TARGET_DATASET --logs-dir $PATH_OF_LOGS

Some examples: ```shell

PersonX -> Market-1501

use all default settings is ok

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainuda.py \ -ds personx -dt market1501 --logs-dir logs/spcluda/personx2marketresnet50

Market-1501 -> MSMT17

use all default settings except for iters=800

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainuda.py --iters 800 \ -ds market1501 -dt msmt17 --logs-dir logs/spcluda/market2msmtresnet50

VehicleID -> VeRi-776

use all default settings except for iters=800, height=224 and width=224

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainuda.py --iters 800 --height 224 --width 224 \ -ds vehicleid -dt veri --logs-dir logs/spcluda/vehicleid2veriresnet50 ```

Unsupervised Learning

To train the model(s) in the paper, run this command:

shell
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python examples/spcl_train_usl.py \
  -d $DATASET --logs-dir $PATH_OF_LOGS

Some examples: ```shell

Market-1501

use all default settings is ok

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainusl.py \ -d market1501 --logs-dir logs/spclusl/marketresnet50

MSMT17

use all default settings except for iters=800

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainusl.py --iters 800 \ -d msmt17 --logs-dir logs/spclusl/msmtresnet50

VeRi-776

use all default settings except for iters=800, height=224 and width=224

CUDAVISIBLEDEVICES=0,1,2,3 \ python examples/spcltrainusl.py --iters 800 --height 224 --width 224 \ -d veri --logs-dir logs/spclusl/veriresnet50 ```

Evaluation

We utilize 1 GTX-1080TI GPU for testing. Note that

• use

  --width 128 --height 256

  (default) for person datasets, and

  --height 224 --width 224

  for vehicle datasets;
• use

  --dsbn

  for domain adaptive models, and add

  --test-source

  if you want to test on the source domain;
• use

  -a resnet50

  (default) for the backbone of ResNet-50, and

  -a resnet_ibn50a

  for the backbone of IBN-ResNet.

Unsupervised Domain Adaptation

To evaluate the domain adaptive model on the target-domain dataset, run:

shell
CUDA_VISIBLE_DEVICES=0 \
python examples/test.py --dsbn \
  -d $DATASET --resume $PATH_OF_MODEL

To evaluate the domain adaptive model on the source-domain dataset, run:

shell
CUDA_VISIBLE_DEVICES=0 \
python examples/test.py --dsbn --test-source \
  -d $DATASET --resume $PATH_OF_MODEL

Some examples: ```shell

Market-1501 -> MSMT17

test on the target domain

CUDAVISIBLEDEVICES=0 \ python examples/test.py --dsbn \ -d msmt17 --resume logs/spcluda/market2msmtresnet50/model_best.pth.tar

test on the source domain

CUDAVISIBLEDEVICES=0 \ python examples/test.py --dsbn --test-source \ -d market1501 --resume logs/spcluda/market2msmtresnet50/model_best.pth.tar ```

Unsupervised Learning

To evaluate the model, run:

shell
CUDA_VISIBLE_DEVICES=0 \
python examples/test.py \
  -d $DATASET --resume $PATH

Some examples: ```shell

Market-1501

CUDAVISIBLEDEVICES=0 \ python examples/test.py \ -d market1501 --resume logs/spclusl/marketresnet50/model_best.pth.tar ```

Trained Models

You can download the above models in the paper from [Google Drive] or [Baidu Yun](password: w3l9).

Citation

If you find this code useful for your research, please cite our paper

@inproceedings{ge2020selfpaced,
    title={Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID},
    author={Yixiao Ge and Feng Zhu and Dapeng Chen and Rui Zhao and Hongsheng Li},
    booktitle={Advances in Neural Information Processing Systems},
    year={2020}
}