Skip to content
A very ugly unofficial implementation of CVPR2016 paper Single-Image Crowd Counting via Multi Column Convolutional Neural Network using tensorflow and keras
Branch: master
Clone or download
Latest commit ca15fda Mar 24, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
keras_modelB
modelA train on 1080 Jan 23, 2018
modelB train on 1080 Jan 23, 2018
samples move img to samples Jan 30, 2018
.gitignore sample images Jan 26, 2018
README.md fix err with heatmap Mar 24, 2018
heatmap.py fix err with heatmap Mar 24, 2018
keras_test.py
keras_train.py
network.py fix err with heatmap Mar 24, 2018
test.py take out mcnn class Jan 30, 2018
train.py take out mcnn class Jan 30, 2018

README.md

Single Image Crowd Counting via Multi Column Convolutional Neural Network

unfinished
unfinished
unfinished

A very ugly unofficial implementation of CVPR2016 paper
Single Image Crowd Counting via Multi Column Convolutional Neural Network
using tensorflow and keras(only for B)

installation

  1. Install tensorflow (and keras)
  2. git clone https://github.com/uestcchicken/MCNN.git

data setup

All the data setup process follows the pytorch version implementation:
svishwa/crowdcount-mcnn

train

For tensorflow:
run python3 train.py A(or B)
model is saved to modelA/ or modelB/

For keras:
run python3 keras_train.py B
model is saved to keras_modelB/

test

For tensorflow:
run python3 test.py A(or B)

For keras:
run python3 keras_test.py B

(uncomment code containing heatmap in network.py to generate heatmap)

result

The model here is trained on DELL laptop, GTX960m, tensorflow-gpu 1.4.1.

A_mae: 119 A_mse: 188
B_mae: 32 B_mse: 55

Using keras:

B_mae: 29 B_mse: 47

In the paper it's:

A_mae: 110 A_mse: 173
B_mae: 26 B_mse: 41

heatmap

actual: 1110

predicted: 1246

You can’t perform that action at this time.