TranSal: Depth-guided Transformer for RGB-D Salient Object Detection

In recent years, RGB-D salient object detection (SOD) has attracted increased attention and has been widely employed in various computer vision applications. Fully convolutional networks dominate many RGB-D SOD tasks and have already achieved outstanding results. However, solving the cross-modality fusion of low-quality depth and RGB cues remains challenging. We present TranSal, an innovative network, depth-guided transformer for RGB-D SOD, based on the Transformer's fantastic performance in image recognition and segmentation. A dual-branch U-Net architecture is used in the proposed model. To begin, it uses ResNet-50 to extract multi-level RGB and depth features. Second, it employs Transformer layers to represent image sequentiality. Finally, the representation is projected into spatial order, and the multi-scale cross-modality characteristics are fused to generate an accurate saliency map using a depth-guided fusion subnetwork. TranSal can successfully mitigate the negative impacts of low-quality depth information and create a saliency map with clear contours and accurate semantics compared to previous models. Experiments and analyses on five large-scale benchmarks verify that TranSal achieves satisfactory performance compared to the recent state-of-the-art methods.