Contour-guided saliency detection with long-range interactions

The guided search theory suggests that global sources of scenes are important for robust visual processing. In this study, we use an eye-tracking experiment to verify that contours play a crucial role in guiding visual attention. Moreover, the experimental data show that subjects pay more attention to the closed regions of line drawings, which are usually related to the dominant objects in the corresponding scenes. In addition, human attention is also guided by scene structure. Inspired by these findings, we designed two long-range selective pooling (LRSP) modules of convolutional neural networks to improve saliency detection by integrating long-range features. The proposed LRSP modules (including regional pooling and layout pooling modules) are inspired by the long-range interactions in the biological visual cortexes and are beneficial for capturing long-range information guided by contours. Extensive experiments show that the proposed method achieves comparable results to state-of-the-art methods, but with a more lightweight network structure. Moreover, benefiting from the contour guidance, the proposed model with long-range interactions demonstrates higher robustness and generalization in the saliency detection task compared with other methods.