Three-branch architecture for stereoscopic 3D salient object detection

Abstract

Existing stereoscopic 3D (S3D) salient object detection (SOD) networks typically employ a two-branch architecture, in which the RGB and depth channels are learned independently. Conventional methods based on conventional neural networks generally fuse the two branches by combining their deep representations at a later stage with only one path, which can be inefficient and insufficient for retaining a large amount of cross-modal data. In this study, we combine the RGB branch and depth branch to generate a third branch. The first branch is the embedded attention branch containing the attention mechanism, and we introduce the embedded attention module in this branch to give the allocation of available processing resources to the most informative components of an input signal. The second branch is the boundary refinement branch combined with the low-level information of RGB and depth images. Additionally, we propose a new module, called the detail correlation module, to ensure clear object boundaries and salient object refinement. The third branch is the global deep-view branch containing the global view module, which fuses high-level information and expands the sensor field. We also use three different loss functions to match our special SOD network. Extensive experiments demonstrate the effectiveness and robustness of the proposed architecture and show that it represents a significant improvement over other state-of-the-art SOD approaches.