Change Detection by Training a Triplet Netowk for Motion Feature Extraction

This repository provides demo code for:

• Training a triplet model to learn motion feature.
• Computing segmentation masks for any image sequence with an extracted MF-Net.

Repository Structure

The repository is structured as follows:

• models contains the Triplet model to train our network
• net contains our trained model
• utils contain lua and CUDA code for our Change Detection algorithm. This needs to be compiled before execution.

Dependencies

This code requires:

• torch7
• opencv >= 2.4(for training)
• png++ (for training)
• cudnn
• display package for torch7

Compute the Segmentation Mask

Download this repository:

$git clone https://github.com/TienPhuocNguyen/mf-net

First, we have to compile the shared libraries:

$cd mf-net/utils
$make

If errors occur, you should check the computing architecture of your GPU and modify the Makefile. Here, we use sm_61 for the Titan X Pascal. If success, the command will produce a file libcutils.so.

To display results on your screen, install the display package:

$luarocks install display

Launch the server:

$th -ldisplay.start 8000 0.0.0.0

Then open 0.0.0.0:8000 on your browser to open the remote desktop.

To execute the program, run the command:

$th run.lua

The file run.lua also provides some arguments to specify the directories of image sequence and trained model. For examples:

$th run.lua -n net/trained.t7 -s datasets/CDNet2014/dataset/dynamicBackground/fall/input