Unsupervised Image Segmentation With GNNs

This repository contains code and data for unsupervised image segmentation using Graph Neural Networks (GNNs). The project includes various datasets, feature extraction methods, and models for segmentation,based on mainly two parer : UNSEGGNET: UNSUPERVISED IMAGE SEGMENTATION USING GRAPH NEURAL NETWORKS and DeepCut: Unsupervised Segmentation using Graph Neural Networks Clustering.

Setup

1. Clone the repository:

   git clone git@github.com:Devnetly/Image_Segmentation_With_GNNs.git
   cd Image_Segmentation_With_GNNs

2. Install the required dependencies:

   pip install -r requirements.txt

Usage

Evaluation

cd src/segmentation
python3 evaluate.py [-h] [--model_name MODEL_NAME] [--feature_type FEATURE_TYPE] \
[--layer LAYER] [--stride STRIDE] [--resize RESIZE] [--segmentation_type SEGMENTATION_TYPE] \
[--threshold THRESHOLD] [--alpha ALPHA] [--activation ACTIVATION] [--num_layers NUM_LAYERS] \ 
[--conv_type CONV_TYPE] [--hidden_dim HIDDEN_DIM] [--num_clusters NUM_CLUSTERS] \ 
[--device DEVICE] [--lr LR] [--n_iters N_ITERS] [--dataset DATASET] [--output_dir OUTPUT_DIR]

Demo

python3 demo.py

Results

ISIC Dataset

Examples :

Results across the different combinations of losses/layers :

Loss	Layer	IoU	Dice	Best Epoch
NCUT	GCN	0.622	0.708	90
NCUT	GAT	0.338	0.409	100
NCUT	ARMA	0.623	0.714	30
CC	GCN (without normalization)	0.434	0.539	20
CC	GAT	0.158	0.219	80
DMON	GCN	0.363	0.483	100
DMON	GAT	0.378	0.498	90
DMON	ARMA	0.500	0.606	80

EMD6 Dataset

Example :

Results across the different combinations of losses/layers :

Loss	Layer	IoU	Dice	Best Epoch
NCUT	GCN	0.642	0.723	50
NCUT	GAT	0.519	0.599	100
NCUT	ARMA	0.649	0.731	30
CC	GCN (without normalization)	0.543	0.636	40
CC	GAT	0.157	0.222	100
DMON	GCN	0.269	0.388	70
DMON	GAT	0.273	0.390	90
DMON	ARMA	0.453	0.558	80