OBFI-dataset

Physically rejoining oracle bone fragments plays an important role in exploring China ancient civilization, a method to accelerate this heavy work effectively is rejoining fragment images, currently this field has received a widespread attention, but it lacks a benchmark dataset. This paper gives a benchmark dataset of oracle bone fragment image including three parts. Part 1, four sub-datasets of single oracle bone fragment image are given and they include 5374 high-resolution images. Part 2 contains four sub-datasets about ground truth rejoinable position of 110 image pairs. Part 3 has two sub-datasets, there are 115,893 unrejoinable and 22,962 rejoinable target area images, which are used to learn texture continuity and test performance in experiment, another 110 target area images of relative complete image and 110 unrejoinable images are used in practical rejoining work. The dataset can be generally used to evaluate rejoining algorithms on various of object fragment images, this field faces two challenges, one is searching matching image edge segments, the other is calculating image texture continuity.