Deep Sort with CSP

CSP version of DeepSORT PyTorch

Project structure

Pretrained model should be copied to here. (Checkpoints for CSP and Deepsort)

./checkpoints

Dataset files in .npz for faster loading

./data_cache

You can put your own dataloader here. (and import them!)

./load_data

Pretrained ResNet50 model should be put here for training

./models

Tasks for project. They also provide visualization.

./tasks

The output directory of tasks. It contains sequence of output images.

./testset_output

Usage

Training:

1. prepare your own dataset or put dataset in data_PETS2009
2. Download pre-trained ResNet50 model. ('https://download.pytorch.org/models/resnet50-19c8e357.pth')
3. Run train_csp.py in terminal or IDE. You can adjust config (eg. image size, batch size, #gpu) in config.py.
4. The checkpoints will be stored in ./weights

Testing

1. Put trained CSP model and Deepsort model checkpoints under ./checkpoints
2. run test_csp.py in terminal or IDE
3. You can also use Tasks.ipynb for evaluation of tasks.

References and Credits:

1. Pytorch implementation of deepsort with Yolo3
2. Center-and-Scale-Prediction-CSP-Pytorch
3. Deep Sort with PyTorch
4. Deepsort
5. SORT
6. PETS2009 Benchmark Data
7. Ground truths for PETS2009 tasks

@inproceedings{liu2018high,
  title={High-level Semantic Feature Detection: A New Perspective for Pedestrian Detection},
  author={Wei Liu, Shengcai Liao, Weiqiang Ren, Weidong Hu, Yinan Yu},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019}
}