The codes are about my paper Adapted Center and Scale Prediction: More stable and More Accurate. This is the pytorch implementation. However, because my paper is based on High-level Semantic Feature Detection: A New Perspective for Pedestrian Detection and this is my first try to begin a computer vision project, I choose to use and change the codes from CSP PyTorch Implementation. Many thanks for his contribution！

Welcome to my website

ACSP PyTorch Implementation

Requirement

• Python 3.6
• Pytorch 0.4.1.post2
• Numpy 1.16.0
• OpenCV 3.4.2
• Torchvision 0.2.0

Reproduction Environment

• Test our models: One GPU with about 4G memory.
• Train new models: Two GPUs with 32G memory per GPU.(If you do not have enough GPU memory resources, please resize the input to 640x1280, it yields slightly worse performance, though.)

Installation

You can directly get the codes by:

  git clone https://github.com/WangWenhao0716/Adapted-Center-and-Scale-Prediction.git

Preparation

1. CityPersons Dataset

You should download the dataset from here. From that link, leftImg8bit_trainvaltest.zip (11GB) is used. We use the training set(2975 images) for training and the validation set(500 images) for test. The data should be stored in ./data/citypersons/images. Annotations have already prepared for you. And the directory structure will be

*DATA_PATH
	*images
		*train
			*aachen
			*bochum
			...
		*val
			*frankfurt
			*lindau
			*munster
		*test
			*berlin
			*bielefeld
			...
	*annotations
		*anno_train.mat
		*anno_val.mat
		...

2. Pretrained Models

The backbone of our ACSP is modified ResNet-101, i.e. replacing all BN layers with SN layers. You can download the pretrained model from here. It is provided by the author of Switchable Normalization. The weights will be stored in ./models/.

3. Our Trained Models

We provide two models:

ACSP(Smooth L1)(code: ydc1): Reasonable 10.0%; Heavy 46.1%; Partial 8.8%; Bare 6.7%.

ACSP(Vanilla L1)(code: 4oa2): Reasonable 9.3%; Heavy 46.3%; Partial 8.7%; Bare 5.6%.

They should be stored in ./models/.

4. Compile Libraries

Before running the codes, you must compile the libraries. The followings should be accomplished in terminal. If you are not sure about what it means, click here may be helpful.

cd util
make all

Training

python train.py

or

CUDA_VISIBLE_DEVICES=x,x python train.py --gpu_id 0 1

Test

python test.py

Citation

If you think our work is useful in your research, please consider citing:

@article{wang2020adapted,
  title={Adapted Center and Scale Prediction: More Stable and More Accurate},
  author={Wang, Wenhao},
  journal={arXiv preprint arXiv:2002.09053},
  year={2020}
}

Note

This is my first computer vision project. If you have any questions or there are something wrong in my codes, feel free to contact me: wangwenhao@buaa.edu.cn.