Unsupervised Segmentation using MASON

Unsupervised Segmentation of Images

Installation

1. Clone this repository

git clone https://github.com/JosephKJ/MASON

Let's call the directory as ROOT

2. Clone the Caffe repository

cd $ROOT
git clone https://github.com/Microsoft/caffe.git

[optional]

cd caffe
git reset --hard 1a2be8e

3. Setup CAFFE

4. Get the ImageNet pretrained VGG-16 network(caffemodel only) and place it in ./model folder.

caffemodel_url: http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel

Demo

You can run python segment_image.py.

This script will take the image located in ./data/demo/2007_000363.jpg, computes the objectness heatmap, uses it with GrabCut and gets the result displayed on screen. (NB: This method doesnot require any bounding box annotation.)

Output (Legend: Input image -- Objectness heat map -- Segmentation using MASON and GrabCut -- Segmentation using GrabCut alone)

semi-supervised-SS.py expects bounding box information from a dataset. The code can work out of the box with Pascal VOC dataset.