DAFNE (Digital Anastylosis of Frescoes challeNgE) Dataset

This dataset was originally created for DAFNE, an international challenge concerned with collecting outstanding contributions to the field of Anastylosis, whose submissions closed in July 2019. The goal was to develop tools enabling digital reconstruction of heavily damaged frescoes, supporting their restoration through the solution of the ‘puzzles’ formed by the remaining fragments, often mixed with spurious elements.
The winners of the challenge presented their results and have been awarded during the 20th International Conference on Image Analysis and Processing (ICIAP 2019) held in Trento, Italy, from 9th to 13th September 2019.

The dataset is now publicly available, but its use is restricted to research, education and non-commercial purposes.

Dataset characteristics

The dataset has been built applying simulated fragmentation to famous frescoes' images. For each fresco you will find: an image of the original fresco and 18 folders, each one containing:

  • a set of fragments — randomly generated — to be reassembled, eventually mixed with spurious elements (PNG files in the folder frag_eroded — each file is a fragment);
  • a list of the co-ordinates of the exact location and orientation of the fragments correctly placed in the reconstructed image (fragments.txt);
  • a list of the spurious fragments (fragments_s.txt);
  • the parameters used for the simulated fresco fragmentation (parameters.txt);
  • the solution consisting in the image of the original fresco with a grey background, with overlapping coloured fragments (rebuilt_image.png).

Example of plane tessellation and ground truth (original image in grey background, with overlapping coloured fragments) for "Cesare Nebbia - The Plague of Milan, 1604"

tessellation01   solution01

The dataset was designed to be realistic, natural and challenging for cultural heritage domains, in terms of image resolution, diversity in scenes, and pictorial assets. The fragments were obtained, from each fresco, through a random plane tessellation with suitable statistics (uniform or with a few clusters as shown in the DAFNE Annex), with an erosion process applied to each fragment.
The data distinguishing characteristics are quantized by the following five basic parameters:

  • A: number of fragments
  • B: type of random distribution
  • C: percentage of missing fragments
  • D: percentage of spurious fragments
  • E: average ratio between the fragment area after the erosion and the original area in the fresco plane tessellation.

The dataset was originally divided in two parts: DB1, provided with all the solutions, to be used as training set for the development phase; and DB2, with no solutions, to be used for the testing challenge phase. Now, the solutions for DB2 (i.e., the exact coordinates and orientation of each fragments, the list of spurious ones, and the reconstructed image) are also provided.

Dataset download

DAFNE dataset download page: to have the credentials to download please send an email to daf-challenge-group@unipv.it specifing in the object "Request for DAFNE dataset access". You will receive a reply as soon as possible.

Citation Information

If you make use of DAFNE dataset, please use the citation for the ICIAP conference and the dedicated “Virtual Special Issue” of Pattern Recognition Letters (Elsevier journal). Please find at the following link the Call for Papers for Special Issue on "Digital Anastylosis of Frescoes challeNgE" (DAFNE).


The images are provided for non-profit purposes and their use is restricted to research, education and non-commercial objectives. Despite every effort to contact copyright-holders, the organizers should be contacted as regards any omissions or errors.