This repository contains the official implementation of our IJCV paper "Collective Migration-Inspired Large-Deformation Compensation for Nonrigid Image Registration".

Please give a star and cite if you find this repo useful.

Inspired by the collective nature of animal migration, we introduce a novel perspective on image registration by framing large-deformation compensation as collective manifold. In this perspective, we regard salient points in images as leaders, similar to leader animals in a migratory flock, playing a pivotal role in guiding the overall registration process. In contrast, non-salient points are considered as flocks, akin to the animals in the flock following the leaders. To estimate a collective manifold, we divide large-deformation registration into three parts following the migration pattern of animals: routes decision based on collectiveness quantification, routes execution with collectiveness maintenance, and cascaded migration with collectiveness inheritance. These three parts are integrated in the collective cascaded migration (CCM) framework, effectively compensating for image large-scale deformation.

• Hybrid Python/MATLAB implementation (requires matlab_engine)

We assess the effectiveness of our method using six diverse datasets. The initial pair comprises 2D semi-synthetic datasets, wherein synthetic deformations are introduced to abdominal slice images. The subsequent four datasets are sourced from challenging open datasets, encompassing three 2D datasets (Fundus Image Registration dataset- FIRE, Object-Kinect- OK, Object-Simulation- OS) and one 3D dataset (Lung CT- LUNG). All datasets exhibit substantial large-scale deformations, visually depicted in the figure below.

The 2D semi-synthetic datasets, NS and NRS, can be downloaded at this link.

Please download the full datasets from their official sources:

• FIRE Dataset
• OK Dataset
• OS Dataset
• LUNG Dataset

Due to GitHub storage limitations, the Data/ folder only contains:

• Subsampled test datasets

conda env create -f ccm.yaml
conda activate ccm

Run the Jupyter notebook files in Scripts/ to test all modules with sample datasets.

If you have any problem, please contact us via driverliu@163.com. We greatly appreciate everyone's feedback and insights. Please do not hesitate to get in touch!

• Pure Python + GPU accelerated code version is coming soon!

Please consider citing our work if you find it useful:

TO BE DONE