A local descriptor based model with visual attention guidance for generic object detection

Attention mechanism of human visual system provides a fast and robust ability to detect objects in cluttered scenes. In this paper, we propose a novel model for generic object detection that combines visual attention guidance and local descriptors representation, without requiring segmentation from background clutter. By matching keypoints of a hierarchical and saliency-based strategy, only the “support” local descriptors are selected to represent the distinctive features of pop-out objects. Simultaneously, the matching threshold is adjusted with saliency weights. Finally, the reference object is located by a simply statistical method among those extracted salient-regions. Two kinds of experiments on sequences and highly cluttered scenes are employed to validate the effectiveness and robustness of the proposed model.