3D Visual Discomfort Prediction Based on Scene Structure and Depth Difference

Abstract

Feature engineering under the stereoscopic images is widely used for visual discomfort prediction. Due to the complexity of the human visual system, extensive feature representation face major challenges, including a lack of inconsistent representation between the response of visual perception and the stimulation of stereoscopic display. To solve this problem, a saliency contrast feature representation method is proposed. Specifically, inspired by human attention mechanism, the stereo image is divided into saliency region and no-saliency region to estimate human fixations from complex natural scenes. The relative distance between saliency region and no-saliency region is obtained to express scene structure features, which combines the visual characteristic and scene structure factor. Besides, depth difference features are represented by combining subjectively perceived depth with objective simulated depth of each region. To evaluate the effectiveness of the proposed method, typical regression models are used to capture the relationship of the extracted features to visual discomfort scores. The experiments conducted on the benchmark IEEE-SA and IVY LAB S3D database show that the superiority of the proposed method and get the optimal solution with random forests as the regression model.