Subjective Evaluation Of Context-Aware Saliency Detection Algorithm

Visual Saliency is the subjective perceptual quality which makes certain parts of an image to stand out more than others. The traditional measurement of visual saliency generally detects the dominant object in the image. A major drawback of this method is that by mainly focusing on the dominant object, its context in the image is lost. The latest saliency detection method – context-aware saliency – detects not only the dominant object but also its surroundings that adds semantic meaning of the scene. In this project, we provide an extensive evaluation of a recently proposed context-aware saliency detection method. The main contributions of this work are in two folds: 1) subjective evaluation framework utilizing EyeLink 1000 eye-tracking system; 2) creation of a data set to provide ground truth data. A representative data set of 60 images was chosen to display to our 17 experiment participants for 4 seconds each. By using the eye tracker, we capture the human gaze pattern needed to understand the context of a scene and use it as ground-truth to evaluate the implemented context-aware saliency detection method. Through comparing the experiment results to the saliency maps created by the algorithm, we identified the strength and the weakness of the algorithm. In addition, we believe that the human fixation data we have collected will be beneficial to the evaluation of various saliency detection methods.