Fig. 1: Qualitative comparison of final saliency output of different methods with our method.
Fig. 2: Precision Recall curve
Fig. 3: Precision recall at maximum F values of different methods.
The standard MSRA 1K Dataset is used in this study which is a subset of widely used MSRA dataset. It contains 1000 images with complex backgrounds and low contrast objects along with their manually labelled ground truth images. For experimental purpose, each input image is divided into regions or superpixels. Each region size is approximately 600 pixels.
In Fig. 1 a qualitative analysis of the study is presented. It is visible that our proposed approach shows better results.
For evaluating performance, precision-recall (PR) curves are used as seen in Fig. 2. Each curve is plotted by comparing the input image’s ground truth against the binary mask of optimized final saliency map which is normalized with a threshold from 0 to 255. Our latest method gives the most appealing result as seen in Fig. 2.
Commonly used PR curves only consider whether the object saliency is higher than the background saliency. So, F-measure is calculated in Fig. 3. Here, also our latest method shows better result than other compared methods.
Here, the compared researches in Fig. 1,2 and 3 are AIM , CA , FT , GB , HC , LC , MSS , SWD , BARC  and BARCS