Evaluation of Segmentation Quality via Adaptive Composition of Reference Segmentations

Bo Peng, Lei Zhang, Xuanqin Mou and Ming-Hsuan Yang








In this work, we build a new image segmentation database and a segmentation evaluation database, respectively. They are free to download for non-commercial research and educational purposes. By downloading these files, you agree not to hold the authors for any damage, lawsuits, or other loss resulting from the possession or use of these images and/or ground truth files. You also acknowledge that you will act according to the terms of use of each image as specified on its source (Corel image library). We reserve the right to change the database at any time without notice. The evaluation functions are given as is without any warranty. The Hong Kong Polytechnic University is not liable for any damage, lawsuits, or other loss resulting from the use of the evaluation functions.

  • Segmentation Database

We establish a new segmentation database which contains 200 source images (from Corel image database), and each image has 6~15 ground truth segmentations.

You can download the database [here] (47.7M).

Sample images and their ground truth segmentations in our segmentation database.


  • Segmentation Evaluation Database

We establish a segmentation evaluation database in order to provide a test bed for segmentation evaluation algorithms. It contains 500 pairs of segmentations (from 500 different images) and their evaluation results by human subjects. The first 200 pairs of segmentations (denoted by Part A) are produced by using the 200 images in our segmentation database, while the other 300 pairs of segmentations (denoted by Part B) are produced by 300 images selected from the Berkeley Image Segmentation Dataset.

You can download the database [here] (30.5M).