Salient Object Detection with Convex Hull Overlap

TL;DR

This paper introduces a novel bottom-up feature called convex hull overlap (CHO) for salient object detection, which enhances traditional appearance contrast features without requiring large-scale training data.

Contribution

It proposes a handcrafted CHO feature that captures spatial overlap cues, filling the gap left by appearance-only methods and eliminating the need for extensive training datasets.

Findings

Achieved positive results on multiple public datasets.

Outperformed some existing methods in salient object detection.

Introduced a new Gestalt-inspired cue for saliency detection.

Abstract

Salient object detection plays an important part in a vision system to detect important regions. Convolutional neural network (CNN) based methods directly train their models with large-scale datasets, but what is the crucial feature for saliency is still a problem. In this paper, we establish a novel bottom-up feature named convex hull overlap (CHO), combining with appearance contrast features, to detect salient objects. CHO feature is a kind of enhanced Gestalt cue. Psychologists believe that surroundedness reflects objects overlap relationship. An object which is on the top of the others is attractive. Our method significantly differs from other earlier works in (1) We set up a hand-crafted feature to detect salient object that our model does not need to be trained by large-scale datasets; (2) Previous works only focus on appearance features, while our CHO feature makes up the gap between the spatial object covering and the object's saliency. Our experiments on a large number of public datasets have obtained very positive results.