Crowd Counting Via Scale-adaptive Convolutional Neural Network

By ZHANG Lu, SHI Miaojing and CHEN Qiaobo
This implementation is written by ZHANG Lu and SHI Miaojing.

Introduction

This project is an implementation of the crowd counting method proposed in our WACV'18 paper - Crowd counting via scale-adaptive convolutional neural network (SaCNN). SaCNN extracts feature maps from multiple layers and adapts them to produce the final density map. A relative count loss is proposed to improve the network generalization on crowd scenes with few pedestrians; a new dataset SmartCity is collected for this scenario.

License

This code is released under the MIT License (Please refer to the LICENSE file for details). It can only be used for academic research purposes. Tencent has all the rights reserved.

Citation

Please cite our paper in your publications if it helps your research:

@article{zhang17sacnn,
Author = {Lu Zhang*, Miaojing Shi* and Qiaobo Chen},
Title = {Crowd Counting Via Sacle-adaptive Convolutional Neural Network},
booktitle= = {IEEE Winter Conference on Applications of Computer Vision (WACV)},
Year = {2018}
}

Dependencies and Installation

We have tested the implementation on Linux with GPU Nvidia Tesla M40. CUDA7.5 and CuDNN v5 is tested. The other version should be working. Caffe installation is pre-required. Inside the repository, lmdb2txt.cpp has to be compiled with caffe under the folder of /yourcaffe/tools/; the euclidean_loss_norm_layers are for the count loss implemtantion in this paper, they should be compiled with caffe as well (caffe.proto is provided).

Training and Test

1. Clone the SaCNN repository

$ git clone https://github.com/miao0913/SaCNN-CrowdCounting-Tencent_Youtu.git

2. Train SaCNN:

$ sh train_sacnn.sh

3. Test SaCNN:

$ sh test_sacnn.sh  
$ MATLAB crowdtest 

Pretrained model on ShanghaiTech PartA and PartB can be downloaded from BaiduYun or GoogleDrive.

SmartCity Dataset

We have collected a new dataset SmartCity in the paper. It consists of 50 images in total collected from ten city scenes including office entrance, sidewalk, atrium, shopping mall etc.. Some examples are shown in Fig. 4 in our arxiv paper. Unlike the existing crowd counting datasets with images of hundreds/thousands of pedestrians and nearly all the images being taken outdoors, SmartCity has few pedestrians in images and consists of both outdoor and indoor scenes: the average number of pedestrians is only 7.4 with minimum being 1 and maximum being 14. We use this set to test the generalization ability of the proposed framework on very sparse crowd scenes.

1. The dataset is available on BaiduYun/GoogleDrive. It is collected by Tencent Youtu Lab. Tecent has the copyright of it. These data can only be used for academic research purposes.
2. Detailed instructions are in the README inside the package. Please cite our WACV'18 paper for reference.

Q&A

Please submit a bug report on the Github site of the project if you run into any problems with the code. For other inquries contact Lu Zhang (zhanlgu330@gmail.com)/Miaojing Shi (miao0913@gmail.com).