A new representation of scene layout improves saliency detection in traffic scenes

Abstract

It has been well established that scene context plays an important role in directing visual attention. A robust representation of scene layout is expected to facilitate further analysis of traffic scenes, especially under challenging visual conditions like nighttime and/or on a small-scale dataset. In this work, a new layout representation for traffic scenes is proposed and applied to the popular visual task of saliency detection. First, a general layout representation for traffic scenes is defined as a combination of an original point (Vanishing Point, VP) and two axes along roadsides. Then, a simple algorithm is proposed to build a robust layout representation for traffic scenes, along with an improved VP detection method. Finally, to verify the contribution of the proposed layout representation, a layout-guided saliency detection framework is proposed to improve existing methods by integrating layout-guided prior learned from human fixations collected with an eye-tracking recorder. Experimental results show that the proposed layout representation can significantly improve the performance of various saliency detection methods including classical bottom-up methods and deep-learning-based methods. Moreover, compared with the deep-learning methods, the layout-guided method has an obvious advantage in terms of robustness when only a small-scale dataset is available and under varying visual scenes.