CRFT (CVPR 2026)

Project Page

Official implementation of CRFT: Consistent-Recurrent Feature Flow Transformer for Cross-Modal Image Registration.

Links

• Code: GitHub Repository
• Paper: arXiv
• Pretrained Weights: Release

Authors

Xuecong Liu, Mengzhu Ding, Zixuan Sun, Zhang Li, Xichao Teng

Overview

Data Preparation

To evaluate/train CRFT, you will need to download the required datasets.

• RoadScene
• OSdataset

You can create symbolic links to wherever the datasets were downloaded in the datasets folder

├── datasets
    ├── os_dataset
        ├── train
           ├── image_pair
           ├── truth_flow
           ├── datum
        ├── test
           ├── image_pair
           ├── truth_flow
           ├── datum
        ├── val
           ├── image_pair
           ├── truth_flow
           ├── datum
     ├── RoadScene
        ├── train
           ├── image_pair
           ├── truth_flow
           ├── datum
        ├── test
           ├── image_pair
           ├── truth_flow
           ├── datum
        ├── val
           ├── image_pair
           ├── truth_flow
           ├── datum

Requirements

conda create --name crft python=3.9.7
conda activate crft
conda install pytorch=2.3.1 torchvision=0.18.1 pytorch-cuda=12.1 matplotlib tensorboard scipy opencv -c pytorch -c nvidia
pip install opencv-python==4.8.0.76
pip install numpy==1.26.4
pip install pytorch-lightning loguru joblib tqdm h5py einops

Training

python train.py

Models

We provide models trained on OSdataset and RoadScene respectively. The default path of the models for evaluation is:

├── checkpoints
    ├── CRFT_OSdataset.ckpt
    ├── CRFT_RoadScene.ckpt 

Test

python test.py 

Citation

@inproceedings{liu2026crft,
  title={CRFT: Consistent-Recurrent Feature Flow Transformer for Cross-Modal Image Registration},
  author={Liu, Xuecong and Ding, Mengzhu and Sun, Zixuan and Li, Zhang and Teng, Xichao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={34784--34794},
  year={2026}
}

crft_gated