Salient Region Proposals for Object Detection

Selective attention helps biological vision allocate resources towards processing visual regions of increased saliency. This evolutionary adaptation allows the human brain to maintain accuracy while minimizing wasted resources. However, current two-stage object detectors expend significant computational resources by exhaustively evaluating every possible spatial location in the visual field. Here, we propose a biologically inspired saliency-guided anchor selection algorithm which generates candidate anchors based on saliency, allowing the network to focus computation on the most potentially informative regions. Our implementation utilizes a biologically-motivated proto-object based saliency model to guide region proposals in a Faster R-CNN framework. We evaluate anchor sparsification on the 2017 MS COCO and the 2012 PASCAL VOC benchmarks, demonstrating that saliency-based approaches reduce anchors by 50% with minimal loss in accuracy, and remain competitive even at 95% sparsification, reducing ≈ 267K anchors down to ≈ 13K for the MS COCO dataset and ≈ 150K anchors to ≈ 8K for PASCAL VOC.