ReBiT-Net: Resource-Efficient Bidirectional Transmission Network for RGB-D Salient Object Detection

Existing artificial neural network-based methodologies for salient object detection in RGB-depth (RGB-D) images typically require significant memory and computation time. In this paper, we propose ReBiT-Net, an novel and resource-efficient network designed to addresses this issue. ReBiT-Net utilizes a mobile network for feature extraction and incorporates depth map quality to regulate the fusion of multi-modal features, resulting in top-to-bottom refinement of salient objects using salient information. Empirical evaluations conducted on five benchmark datasets demonstrate the superior performance of our model in terms of accuracy (achieving 334 frames per second for an input size of 320 x\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 320) and model parameters (merely 5.1 MB). Moreover, we introduce ReBiT-Net*, a simplified variant of ReBiT-Net, which entails reduced model parameters (4.2 MB) and enhanced processing speed (793 frames per second for a 256 x\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 256 input size). These improvements are achieved through reduced memory requirements and computational demands via the adoption of a smaller input image size.