This dataset is a collection of salient object boundaries based on Berkeley Segmentation Dataset (BSD). Seven objects are asked to choose the salient object(s) in each image used in BSD. Each subject is shown randomly a subset of the Berkeley segmentation dataset as boundaries overlapped on the corresponding images. He can then choose which regions or segments correspond to salient objects by clicking on them.
For each image of the 300 images used in BSD, there is a .mat file which can be opened by Matlab. Loading each mat file reads a structure called 'SES' into memory which is an array of data collected from sessions each subject in SOD has worked on.
Each element of SES, for example SES(i) has the following structure:
|.session||A unique session number|
|.sj||Subject number (1 to 7)|
|.index||A number identifying which segmentation in BSD was used in this session|
|.obj||An array of structures about objects identified in this session|
|.ImSize||Image Size [height, width]|
|.PERS||This used to refer to subjects' name, but now has subject number|
Each object in obj structure, for example obj(j) contains the following information:
|.BND||This is a cell array, where each cell contains a nx2 matrix of coordiantes of n points on the boundary of the object. If the object has non-adjacent parts, there will be more than one cell in bnd, each corresponding to a separate part.|
|.IMP||Importance or salience of object. The most salient object has imp=1.|
|.seg||An array indicating the segments (in BSD) constructing the object.|
|.objID||A number assigned to the object (1 to number of objects identified by all subjects in all sessions)|
|.MMC||A confidence value indicating consistency with objects identified by other subjects.|
The confidence value is calculated as follows:
For each image in SOD, there are 7 subjects. Suppose subject j has identified object O in SOD. To assign a confidence measure to this object, it's consistency with objects identified by other subjects should be considered. If the maximum consistency between this object and all objects of subject k is denoted as C(k), then MMC is the average of C(k) for all six other subjects:
MMC(O,j)= mean(C(k)), k=1..7, k not equal to j
A threshold, e.g. 0.9, can be used to filter out objects with less confidence values.
Consistency is measured as the intersection to union ratio for two objects.
The name of each file in SOD is formatted as "SOnnnn.mat", wher nnnn refers to the image number (not necessarily 4 digits). DBidx.mat has summarized information on SOD, which might be useful in some applications:
|SFprefix||An array of 300 image numbers (nnnn in SOD file names)|
|MAXSES||The ith element of this array, contains the number of seesions available for image i (i=1..300). This equals to the length of 'SES' strucure if the corresponding SOD file is read.|
|MAXOBJ||The (i,j)th element of this matrix shows the number of objects identified in session j of image i. This is equal to the length of 'obj' in SES(j) of SOD file corresponding to image i.|
Note that the original images are available from the Berkely Segmentation Dataset at:
Please cite our paper: Vida Movahedi and James H. Elder (2010), "Design and Perceptual Validation of Performance Measures for Salient Object Segmentation",
in 7th IEEE Computer Society Workshop on Perceptual Organization in Computer Vision (POCV), San Francisco, CA.
Download SOD(11.0 MB)
Download Matlab code to generate binary images of object boundaries in SOD
Back to main page