ling-bao / acscp_cgan Goto Github PK

Code implementation for paper that "ACSCS: Crowd Counting via Adversarial Cross-Scale Consistency Pursuit"; This is method of Crowd counting by conditional generation adversarial networks

License: GNU General Public License v3.0

Python 61.38% MATLAB 38.62%

crowd-counting deep-learning conditional-gan acscp

acscp_cgan's Introduction

HEAD

ACSCP crowd counting model

=======

Introduction

This is open source project for crowd counting. Implement with paper "Crowd Counting via Adversarial Cross-Scale Consistency Pursuit" from Shanghai Jiao Tong University. For more details, please refer to our Baidu Yun

Installation
Preparation
Train/Eval/Release
Additional
Details

Installation

Configuration requirements

python3.x

Please using GPU, suggestion more than GTX960

python-opencv
#tensorflow-gpu==1.0.0
#tensorflow==1.0.0
scipy==1.0.1
matplotlib==2.2.2
numpy==1.14.2

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

Get the code

git clone [email protected]:Ling-Bao/ACSCP_cGAN.git
cd ACSCP_cGAN

Preparation

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.
Get dataset and its corresponding map label Baidu Yun Password: yvs1
Unzip dataset to ACSCP_cGAN root directory

unzip Data.zip

Train/Eval/Release

Train is easy, just using following step.

Train. Using main.py to train crowd counting model

python main.py --phase train

Eval. Using main.py to evalute crowd counting model

python main.py --phase test

OR

python main.py --phase inference

Model release Model release. Using product.py to release crowd counting model. Download release version 1.0.0, please click on here

Addtional

Crowd map generation tools Source code store in "data_maker", detail please check here. **Note: **This tools write by matlab, please install matlab.
Results

Original image

Real crowd map, counting is 707

Predict crowd map, counting is 698

crowd counting paper collection, thanks for gjy3035 Github: Awesome-Crowd-Counting Density Map Generation from Key Points: [Matlab Code] [Python Code]

Details

Tring to delete dropout layers.

=======

TAIL

acscp_cgan's People

Contributors

Stargazers

Watchers

Forkers

seeker1943 dtn97 anzhicheng tlwzzy liuzhanfeng123 sijar-1 whalesi blesszd wadx2019 stc-cqupt xinxinatg fasladodo michael-d-666 jinbang129 bayanalmutawa ajitaru

acscp_cgan's Issues

tensorflow版本问题

ValueError: Variable vgg_16/conv1/conv1_1/weights already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?
出现这个问题，看前面的问答，是版本问题，请问如何修改代码以适应高版本呢？

标注问题

为什么我进行标注完成之后生成GT_mat文件没有后缀并且标注一次覆盖一次有谁能解决一下么？

ROI 问题

您好，不知道您在UCSD和worldexpo'10这两个数据集做过测试没有，这两个数据集都有一个ROI区域的限制。那最后对密度图求和的时候，是采用整张图求和还是对密度图的ROI部分进行求和？

论文的例子复现问题

您好，我尝试着跑了一些论文出现的例子，比如ＳｈａｎｇｅｈａｉＴｅｃｈ/part_B/test_data/images/IMG_141.jpg，（论文中Ｆｉｇｕｒｅ 4.的第５行），得到的ｅｓｔｉｍａｔｅd=236, 与论文中的３３．９相差甚远。

我是参考的ｐｒｏｄｕｃｔ.py的代码去跑的，　图片ｒｅｓｉｚｅ成７２０ｘ７２０。请问我使用的方法对吗？

提前感谢。

关于代码运行报错

您好！很高兴您能够分享您的项目，不过我用您的代码直接运行python main.py --phase train会报错，提示：
ValueError: Variable vgg_16/conv1/conv1_1/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:
File "/home/user/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 216, in variable
use_resource=use_resource)
File "/home/user/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 181, in func_with_args
return func(*args, **current_args)
File "/home/user/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 261, in model_variable
use_resource=use_resource)
即使将vgg_feature.py中的相关函数注释掉也还是一样的，请问您能帮忙解释下么？跪求！

Fixed map_mcnn.m for creating density maps (Step 3)

For anyone having trouble creating density maps using the given map_mcnn.m, try using this one which I made a few changes to.
MCNN_Density_Matlab.zip

关于ACSCP算法实现的几个疑问，还请指教……

您好！很高兴您能够分享关于ACSCP算法的实现，关于您的实现我有几个疑问还请指教：
1）论文中说，生成网络的最后一层后面应该是tanh激活函数，但是看您用的是先Sigmoid后ReLU？
2）生成网络是一个U-Net的结构，在镜像对称的卷积和反卷积层中间添加了skip connections，不应该是e1连接d8、e2连接d7吗？看您是e1连接d7、e2连接d6呢？
3）训练方面，作者训练了300个epoch，前100个epoch，λc=0；后200个epoch λc=10，不知您是否是这样训练的？
4）另外您给出的结果示例，应该是IMG_173_A的图吧，可是原测试数据集中给的ground truth是483个人，而您给出的结果中ground truth人数是707人呢？
5）我用您的代码和数据集训练，在product时生成的密度图咋有很多竖条纹？
6）由密度图得出人数，直接round(sum(sum(mp)))就可以了，这是什么原理？
希望您能够在百忙之中抽出时间予以解答，不生感激，谢谢！

Thanks.