SBCoseg Dataset: A New Public Image Dataset Characterized by Simple Background

    Abstract

    We established a new co-segmentation dataset (SBCoseg), which contains 66 image groups with 18 images/group and pixel-wise hand-annotated ground truth. The images in SBCoseg are characterized by simple backgrounds produced from a single color. However, it is still difficult for current co-segmentation algorithms. The difficulties mainly exist in shadow removal, holes removal, the discrimination between foreground and easy confused background, and the color restoration of transparent regions. We think that such relatively simpler co-segmentation problem should be solved firstly before that for complicated background can be solved. Our original established SBCoseg dataset can serve as a new starting point for solving the co-segmentation problem.

    Introduction

    Over the last decades, a lot of co-segmentation algorithms have been proposed. A comprehensive evaluation of these algorithms can provide an understanding of the state-of-the-art and help to analyze the strengths and the weaknesses of each algorithm. For this purpose, evaluation datasets are needed. The commonly used datasets for evaluating co-segmentation algorithms include iCoseg, MSRC, ImageNet, and FilckerMFC. The image backgrounds in these datasets are all complicated, which vary in color or intensity. The exact co-segmentation over these images with complicated background is difficult to be obtained for current algorithms. Thus, we establish a dataset of images with simple background and use it to test and analyze the current co-segmentation algorithms.

    Dataset Description

    The SBCoseg dataset includes 66 groups of images and each group consists of 18 images with common object, leading to 1188 images in total. Each image is in JPG format with pixel size of 720720. Some image examples and their corresponding ground truths in SBCoseg are illustrated in the following figure.

    

    Dataset Challenges

    1) The shadow of an object can be easily labeled as foreground, especially for the shadows next to the object.

    2) The images with easy confused foreground and background are hard to be exactly segmented.

    3) The displaying color of the transparent regions are merged with the background color, and to restore their color is difficult.

    4) The holes, especially minor holes in the object are hard to be removed.

    Rules

    The SBCoseg dataset is established for the development of co-segmentation. We do not want to create any obstacles for publishing methods that use this dataset. At the same time, we ask the people who download or use the SBCoseg to abide the rules below:

    1) Data downloaded from this site may only be used for the purpose of scientific studies, for example, may be used to train or develop new algorithms for scientific studies. All the data from SBCoseg cannot be used to train or develop algorithms used in commercial products.

    2) When the data and/or the results of algorithms on the data are used in scientific publications (journal publications, conference papers, technical reports, presentations at conferences and meetings) you must state the data sources.

    3) Teams must notify the maintainers of this site about any publication that is (partly) based on the data on this site, in order for us to maintain a list of publications associated with the SBCoseg dataset.

    Download

    You can download SBCoseg dataset from this link:   http://www.iscbit.org/source/MLMRdataset.zip. 

     Notice that you are not allowed to download these data if you do not agree with the above rules.

    Contact

    If you are not yet clear about the dataset, or if you have additional questions, please e-mail Mengqiao Yu (yumengqiao@bit.edu.cn).

    Address: Machine Learning and Muti-Media Retrieval Lab, Beijing Institute of Technology, Beijing 100081, China.

 

Address: Room 901,Central Teaching Building,Beijing Institute of Technology, Beijing, P. R.China.

TEL: +86-010-68913447