In this page a dataset for retinal image registration, annotated with ground truth data, is availed.

Please cite the following paper if utilizing any part of the dataset:

FIRE: Fundus Image Registration Dataset
C. Hernandez-Matas, X. Zabulis, A. Triantafyllou, P. Anyfanti, S. Douma, A.A. Argyros
Journal for Modeling in Ophthalmology, vol. 1, no. 4, pp. 16-28, Jul. 2017.

Dataset Contents

The dataset consists of 129 retinal images forming 134 image pairs. These image pairs are split into 3 different categories depending on their characteristics. The images were acquired with a Nidek AFC-210 fundus camera, which acquires images with a resolution of 2912x2912 pixels and a FOV of 45° both in the x and y dimensions. Images were acquired at the Papageorgiou Hospital, Aristotle University of Thessaloniki, Thessaloniki from 39 patients.

Download FIRE Dataset (264 MB)

The file contains:


The images follow the naming convention:
[Image pair name]_X.jpg
where X is 1 for the reference image and 2 for the test image.

Ground truth

The ground truth files follow the naming convention:
control_points_[Image pair name]_1_2.txt

The ground truth file for each image pair has the following format:
[reference_point_1_x] [reference_point_1_y] [test_point_1_x] [test_point_1_y]
[reference_point_2_x] [reference_point_2_y] [test_point_2_x] [test_point_2_y]

Registration executable

To run experiments on this dataset, the Registration through Eye Modelling and Pose Estimation (REMPE) retinal image registration method has been publicly available at


For any issue regarding this page or dataset, please contact: Carlos Hernandez-Matas

Registration Scores

Please cite the corresponding papers from which you use the results. If you wish to add your results here, please contact us.

Scores from diverse publications are presented here. Error for each image pair is calculated as the mean registration error for each image pair.

The x axis of the plot corresponds to the value of an error threshold. If the registration error of an image pair is below this threshold, the registration is considered as successful. The y axis of the plot corresponds to the percentage of successfully registered image pairs for a given threshold.

The table indicates the Area Under Curve (AUC) for the plots, facilitating comparison when methods perform similarly. Higher is better.

1 REMPE (H-M 17) 0.958 0.542 0.660 0.773
2 H-M 16 0.945 0.443 0.577 0.721
3 Harris-PIIFD 0.900 0.090 0.443 0.553
4 GDB-ICP 0.814 0.303 0.303 0.576
  1. REMPE (H-M 17)
    REMPE: Registration of Retinal Images through Eye Modelling and Pose Estimation
    C. Hernandez-Matas, X. Zabulis, A.A. Argyros
    IEEE Journal of Biomedical and Health Informatics, 2020
    DOI: 10.1109/JBHI.2020.2984483
    Download scores.
  2. H-M 16
    Retinal Image Registration Through Simultaneous Camera Pose and Eye Shape Estimation
    C. Hernandez-Matas, X. Zabulis, A.A. Argyros
    38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 3247-3251, Orlando, August 16-20, 2016
    DOI: 10.1109/EMBC.2016.7591421
    Download scores.
  3. Harris-PIIFD
    A Partial Intensity Invariant Feature Descriptor for Multimodal Retinal Image Registration
    J. Chen, J. Tian, N. Lee, J. Zheng, R. T. Smith, A. F. Lane,
    IEEE Transactions on Biomedical Engineering, vol. 57, no. 7, pp. 1707-1718, Feb. 2010
    DOI: 10.1109/TBME.2010.2042169
    Download scores.
  4. GDB-ICP
    Registration of Challenging Image Pairs: Initialization, Estimation, and Decision
    G. Yang, C. V. Stewart, M. Sofka, C. L. Tsai
    IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 29, no. 11, pp. 1973-1989, Nov. 2007
    DOI: 10.1109/TPAMI.2007.1116
    Download scores.


This research was made possible by a Marie Curie grant from the European Commission in the framework of the REVAMMAD ITN (Initial Training Research Network), Project 316990. It was also supported by the FORTH-ICS internal RTD Programme "Ambient Intelligence and Smart Environments".