CSP-pedestrian-pytorch

PyTorch implementation of CSP [https://github.com/liuwei16/CSP]

This code is only for Caltech dataset currently, and only for center-position+height regression+offset regression model.

We will add Citypersons dataset support in the future.

Note

On Caltech validation set, we get the best result is 5.84 MR.

Dependencies

• Python 3.7
• PyTorch 1.5.1 + torchvision 0.6.1
• OpenCV 4.3.0.36
• MMCV 0.6.2

Installation

1. Get the code.

  git clone https://github.com/polariseee/CSP-pedestrian-pytorch.git

2. Compile NMS.

  cd ./external
  python setup.py build_ext --inplace
  rm -rf ./build

Preparation

1. Download the dataset.

For pedestrian detection, you should firstly download the datasets. For Caltech, we assume the dataset is stored in ./Caltech/.

2. Dataset preparation.

For Caltech, the directory structure is

*DATA_PATH
   *train
       *IMG
           *set00_V000_I00002.jpg
           *...
       *anno_train10x_alignedby_RotatedFilters
           *set00_V000_I00002.txt
           *...
   *test
       *IMG
           *set06_V000_I00029.jpg
           *...
       * anno_test_1xnew
           *set06_V000_I00029.jpg.txt
           *...

Training

1. Train on Caltech

  python train.py config/config.py

note: If you use one gpu, please modify the parameter chunk_sizes in config/config.py.

Test

1. Caltech

  python test.py config/config.py

Evaluation

1. Caltech

You should use matlab to evaluate your results.Meantime, Caltech toolbox should be download from official website, and the toolbox is sorted in ./eval_caltech/toolbox.

Follow the ./eval_caltech/dbEval.m to get the Miss Rates of detections