An investigation of gradient as a feature cue for saliency detection

Salient object detection is a prominent research topic, based on a human's ability to selectively process conspicuous objects/regions within a scene. With many low-level features being adopted into saliency models, gradient is often overlooked. We investigate the effectiveness of gradient as a feature, applying and evaluating multiple image gradient operators. Scale is also addressed via the use of different sizes of convolutional masks and by varying the neighbour region to calculate gradient contrast. Finally, we present and evaluate a single scale saliency model with the respective gradient cue from each operator, for the detection of salient objects. Each model is evaluated on the publicly available MSRA10K salient object dataset.