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