This repository includes the official implementation of LibraNet for crowd counting, presented in our paper:

Weighing Counts: Sequential Crowd Counting by Reinforcement Learning

Proc. European Conference on Computer Vision (ECCV), 2020

Liang Liu^1,†, Hao Lu^2,†, Hongwei Zou¹, Haipeng Xiong¹, Zhiguo Cao¹, Chunhua Shen¹

¹Huazhong University of Science and Technology

²The University of Adelaide, Australia

^† equal contribution

The code has been tested on Python 3.7.6 and PyTorch 1.4.0. Please follow the official instructions to configure your environment. See other required packages in requirements.txt.

• Download the pre-processed ShanghaiTech Part_A training set from: BaiduYun (168.3 MB) (code: ix2v) or OneDrive (172.3 MB).

• Download the ShanghaiTech Part_A testing set from: BaiduYun (23.7 MB) (code: h7a6) or OneDrive (24.3 MB).

• Unzip the datasets and move 'Train' and 'Test' folder into './data', the path structure should look like this:

$./data/
├──── Train
├──── Test

• Download the VGG16 backbone pretrained on SHT Part_A from [BaiduYun (56.1 MB) (code: 3cfp) or OneDrive (57.5 MB)](https://1drv.ms/u/s!AkNf_IPSDakh8jLP6doilJNgdr4g?e=JcgOMV).

• Move the backbone model into the folder, and the path structure should like this::

$./backbone.pth.tar

Train LibraNet on SHT Part_A Dataset

python train.py

Pre-trained Model on SHT Part_A dataset

• Download the model from: [BaiduYun (68.3 MB) (code: 20um) or OneDrive (70 MB)](https://1drv.ms/u/s!AkNf_IPSDakh8XBVTepnGq2J_YjN?e=lJCCUw)
• The result of this model is: mae=55.5, mse=93.9. However, if the pythorch version is less than 1.4.0 (1.3.1 for example), the result might be: mae=56.3 , mse=95.2. Now I try to find the reason.
• Move the model into the folder, and the path structure should like this:

$./trained_model/
├──── LibraNet_SHT_A.pth.tar

Evaluation

python Test_SHT_A.py

If you find this work or code useful for your research, please cite:

@article{liu2020WeighingCounts,
  title={Weighing Counts: Sequential Crowd Counting by Reinforcement Learning},
  author={Liu, Liang and Lu, Hao and Zou, Hongwei and Xiong, Haipeng and Cao, Zhiguo and Chun, Huashen},
  journal={Proc. Eur. Conf. Computer Vision},
  year={2020}
}

1. Fix a bug in train_test.py line 32

• Error:

for image_index in range(0, 1):

• Correct:

for image_index in range(0, train_number):

2. Add LICENSE.md

The code are only for non-commercial purposes. Copyrights reserved.

Contact: Liang Liu (wings@hust.edu.cn) Hao Lu (poppinace@foxmail.com)