search

Ling-Bao / mscnn

mscnn crowd counting model implementation, source from "Multi-scale Convolution Neural Networks for Crowd Counting" write by Zeng L, Xu X, Cai B, et al.
https://arxiv.org/abs/1702.02359
GNU General Public License v3.0
93 stars 36 forks source link
crowd-counting deep-learning

readme

HEAD

mscnn crowd counting model

======= License

Introduction

This is open source project for crowd counting. Implement with paper "Multi-scale Convolution Neural Networks for Crowd Counting" write by Zeng L, Xu X, Cai B, et al. For more details, please refer to arXiv paper

multi-scale block

mscnn_model

mscnn_architecture

result_display

result_table

Contents

  1. Installation
  2. Preparation
  3. Train/Eval
  4. Details

Installation

  1. Configuration requirements 
    
python3.x

Please using GPU, suggestion more than GTX960

python-opencv

tensorflow-gpu==1.0.0

tensorflow==1.0.0

matplotlib==2.2.2 numpy==1.14.2

conda install -c https://conda.binstar.org/menpo opencv3 pip install -r requirements.txt

2. Get the code

git clone https://github.com/Ling-Bao/mscnn cd mscnn

### Preparation
1. ShanghaiTech Dataset. 
ShanghaiTech Dataset makes by Zhang Y, Zhou D, Chen S, et al. For more detail, please refer to paper "Single-Image Crowd Counting via Multi-Column Convolutional Neural Network" and click on [here](https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Zhang_Single-Image_Crowd_Counting_CVPR_2016_paper.pdf).

2. Get dataset and its corresponding map label
[Baidu Yun](https://pan.baidu.com/s/12EqB1XDyFBB0kyinMA7Pqw) 
Password: sags

3. Unzip dataset to mscnn root directory

tar -xzvf Data_original.tar.gz


### Train/Eval
Train is easy, just using following step.

1. Train. Using [mscnn_train.py](mscnn_train.py) to evalute mscnn model

python mscnn_train.py


2. Eval. Using [mscnn_eval.py](mscnn_eval.py) to evalute mscnn model

python mscnn_eval.py



### Details
1. Improving model structure. Add Batch Normal after each convolution layer.

=======
[![License](http://gplv3.fsf.org/gplv3-127x51.png)](LICENSE)
>>>>>>> TAIL