Channel-overcomplete convolutional architectures for RGB-D salient object detection

Abstract

It is a challenging task to detect small salient objects by fusing color information and spatial information from RGB-D images. Most RGB-D salient object detection (SOD) methods use U-Net or its variants because they tend to extract high-level features that can quickly determine the location of salient objects. However, it is also indispensable to effectively learn low-level features to detect fine boundaries and small structures. To this end, we propose a novel two-branch SOD network (COSU), which includes channel-overcomplete (CO) and spatial-undercomplete (SU) branches. First of all, the CO branch pays attention to low-level features by transform the spatial distribution into the channel connection and restricting the receptive field of filters. Secondly, SU branch uses increased receptive field to capture the high-level features. Finally, a multi-scale modal correction strategy is designed to fully fuse the multi-level features from the two branches and mitigate the modal divergence from RGB-D features. Experimental results on seven widely used datasets show that the proposed COSU outperforms ten state-of-the-art RGB-D SOD methods.