Arbicon-Net (Arbitrary Continuous Geometric Transformation Networks for Image Registration)

Arbicon-Net is initially described in a NeurIPS 2019 paper.

Dependency

The code is implemented using Python 3 and PyTorch 0.2. A quick installation in Anaconda:

conda env create -f environment.yml

Getting Started

Training

To train CNNGeo-aff in synthetic dataset please run:

cd scripts/
bash train_strong_random_affine_pascal.sh

To train Arbicon-Net in synthetic dataset please run:

cd scripts/
bash train_strong_random_tps_pascal.sh

To fine-tune the two-stage network in PF-PASCAL training set please run:

cd scripts/
bash train_weak_pf_pascal.sh

Evaluation

To evaluate your model please run:

python eval.py --feature-extraction resnet101 --model [your model] --eval-dataset [evaluation dataset]

To further evaluate your model on TSS dataset please run utils/tss_eval/Main.m in MATLAB. Trained weight could be found and downloaded here.

Main Results

PF-PASCAL:

Method	PCK
WeakAlign	75.8
Arbicon-Net	77.3

PF-Willow:

Method	PCK
WeakAlign	71.2
Arbicon-Net	72.2

TSS:

Method	FG3D	JODS	PASC
WeakAlign	90.3	76.4	56.5
Arbicon-Net	92.5	76.5	58.5

Acknowledgement

• [weakalign], [NC-Net] by Ignacio Rocco et al.
• [TSS Evaluation Toolkit] by Tatsunori Taniai et al.

Bibtex

If you find our paper helpful, please cite the paper.

@InProceedings{chen2019nips,
author = {Jianchun Chen, Lingjing Wang, Xiang Li, and Yi Fang},
title = {Arbitrary Continuous Geometric Transformation Networks for Image Registration},
booktitle = {Proceedings of the Neural Information Processing Systems (NeurIPS 2019)},
year = {2019}
}

This code is provided for academic use only. For any question please contact Jianchun Chen ([email protected]).