RGB-Guided Depth Feature Enhancement for RGB–Depth Salient Object Detection

Abstract

RGB-D (depth) Salient Object Detection (SOD) seeks to identify and segment the most visually compelling objects within a given scene. Depth data, known for their strong discriminative capability in spatial localization, provide an advantage in achieving accurate RGB-D SOD. However, recent research in this field has encountered significant challenges due to the poor visual qualities and disturbing cues in raw depth maps. This issue results in indistinct or ambiguous depth features, which consequently weaken the performance of RGB-D SOD. To address this problem, we propose a novel pseudo depth feature generation-based RGB-D SOD Network, named PDFNet, which can generate some new and more distinctive pseudo depth features as an extra supplement source to enhance the raw depth features. Specifically, we first introduce an RGB-guided pseudo depth feature generation subnet to synthesize more distinctive pseudo depth features for raw depth feature enhancement, since the discriminative power of depth features plays a pivotal role in providing effective contour and spatial cues. Then, we propose a cross-modal fusion mamba (CFM) to effectively merge RGB features, raw depth features, and generated pseudo depth features. We adopt a channel selection strategy within the CFM module to align the pseudo depth features with raw depth features, thereby enhancing the depth features. We test the proposed PDFNet on six commonly used RGB-D SOD benchmark datasets. Extensive experimental results validate that the proposed approach achieves superior performance. For example, compared to the previous cutting-edge method, AirSOD, our method improves the F-measure by 2%, 1.7%, 1.1%, and 2.2% on the STERE, DUTLF-D, NLPR, and NJU2K datasets, respectively.