DynaSeg: A deep dynamic fusion method for unsupervised image segmentation incorporating feature similarity and spatial continuity

Authors: Boujemaa Guermazi, Riadh Ksantini, Naimul Khan

Official repository for DynaSeg: A Deep Dynamic Fusion Method for Unsupervised Image Segmentation. This includes clean code and instructions for reproducing the results reported in our Image and Vision Computing journal paper.

Overview

DynaSeg is a deep dynamic fusion method for unsupervised image segmentation that effectively balances feature similarity and spatial continuity. It automates parameter tuning through dynamic weighting, allowing for efficient segmentation without extensive hyperparameter adjustments. DynaSeg integrates seamlessly with existing segmentation networks and has been tested to achieve state-of-the-art performance on multiple benchmark datasets.

Alt text

Installation

To set up DynaSeg, you need to install the following OpenMMLab packages:

MIM >= 0.1.5
MMCV-full >= v1.3.14
MMDetection
MMSegmentation
MMSelfSup

Install the required packages using the following commands:

pip install openmim mmdet mmsegmentation mmselfsup
mim install mmcv-full

Data Preparation

To prepare the necessary data for training, follow these steps:

Download Datasets:
- Training Set: Download the training set of the COCO dataset.
- Validation Set: Download the validation set of the COCO dataset.
- Stuffthing Map: Download the stuffthing map.
Unzip and Organize Data: Unzip the data and organize it into the following directory structure:

data/
├── curated
│   ├── train2017
│   │   ├── Coco164kFull_Stuff_Coarse_7.txt
│   ├── val2017
│   │   ├── Coco164kFull_Stuff_Coarse_7.txt
├── coco
│   ├── annotations
│   │   ├── train2017
│   │   │   ├── xxxxxxxxx.png
│   │   ├── val2017
│   │   │   ├── xxxxxxxxx.png
│   ├── train2017
│   │   ├── xxxxxxxxx.jpeg
│   ├── val2017
│   │   ├── xxxxxxxxx.jpeg

The curated directory contains the data splits for unsupervised segmentation, structured according to the splits used by PiCIE.

Testing

To evaluate DynaSeg on different datasets, follow these instructions:

COCO Dataset:

Use the script tools/DynaSeg_COCO.py to run the testing on the COCO dataset.
The mIoU (mean Intersection over Union) for all images will be stored in eval_results.pkl.

Pascal VOC 2012 and BSD500 Datasets:

Use the script tools/DynaSeg_SCF_BSD_VOC.py to run the testing on the Pascal VOC 2012 and BSD500 datasets.
After running the tests, you can calculate the mIoU using the script tools/calculate_mIOU.py.

Acknowledgement

This codebase is based on MMCV, and it benefits a lot from DenseSiam, PiCIE, MMSelfSup, and Detectron2.

License

This project is licensed under the MIT License.

Citation

If you use DynaSeg in your research, please cite our work:

@article{guermazi2024dynaseg,
  title={DynaSeg: A deep dynamic fusion method for unsupervised image segmentation incorporating feature similarity and spatial continuity},
  author={Guermazi, Boujemaa and Ksantini, Riadh and Khan, Naimul},
  journal={Image and Vision Computing},
  pages={105206},
  year={2024},
  publisher={Elsevier}
}

RyersonMultimediaLab / DynaSeg

readme