The TUM-Image Inpainting Database


Introduction to the TUM-Image Inpainting Database

The availability of digital photography has increased in recent years. Nowadays, most mobile phones are equipped with cameras. Naturally, the interest in image manipulation tools is growing. A common demand is the removal of interfering objects or persons from an image. Obviously, the removal of parts of the image leaves holes which need to be filled. This task is called image inpainting.
When presenting a new inpainting algorithm the authors show usually side by side comparisons between their inpainted images and results of existing algorithms. The problem is that no database of inpainted images with including baseline images exists. Only with a common database wider comparison and statements can be made.

Dataset Description and Contribution

We provide a database consisting of 17 base images. These images differ in relation to texture and structure diversity. Each of these images is inpainted by four state-of-the-art inpainting algorithms [1-4].
We use two compact and two spread-out target regions, which define the inpainting area. The compact region is a blob-like area, whereas the spread-out region consists of thinner lines distributed over a wider space in the image. The occupied size in the image, however, is in both cases 5% for the small and 10% for the big-sized target region.
In total 272 inpainted images are available which allow comprehensive comparison between new and existing algorithms. The template of the target regions are available too.

[1] Aurélie Bugeau, Marcelo Bertalm´ io, Vicent Caselles, and Guillermo Sapiro, “A comprehensive framework for image inpainting,” IEEE Transactions on Image Processing, vol. 19,no. 10, pp. 2634–2645, 2010.

[2] Jan Herling and Wolfgang Broll, “Pixmix: A real-time approach to high-quality diminished reality,” in International Symposium on Mixed and Augmented Reality. 2012, pp. 141–150, IEEE

[3] Pascal Getreuer, “Total variation inpainting using split bregman,” Image Processing On Line, vol. 2, pp. 147–157, 2012.

[4] Zongben Xu and Jian Sun, “Image inpainting by patch propagation using patch sparsity,” IEEE Transactions on Image Processing, vol. 19, no. 5, pp. 1153–1165, 2010


The dataset is provided as one compressed and password protected zip file. The password is mmk@inpainting.

Folder Structure

Three folders exists inside the zip-file. The "images" folder includes all base images. The "inpaintedImages" folder holds all inpainted images. The names of those images are following the naming scheme below. The last folder "masks" includes the four masks in "png" and "svg" file format.

Naming Convention

The inpainted images files are named "algorithm-ImageNr-TargetRegionNr.png", where algorithm is one of the four author names, ImageNr is a number between 1-17 referring to the base images (see folder "images"). TargetRegionNr defines the kind of applied target region (see folder "masks").

TUM-IID Download


If you use this database in your research, please cite the following paper:

P. Tiefenbacher, V. Bogischef, D. Merget and G. Rigoll, “Subjective and Objective Evaluation of Image Inpainting Quality”, in Proc. ICIP, IEEE, 2015


Please contact Philipp Tiefenbacher for questions and the like.