GAN-Based RGB-to-Depth Generation for Enhanced Salient Object Detection
Sudipta Bhuyan · 2024
Depth information is crucial for detecting salient objects in intricate scenarios, leading to the rising popularity of RGB-D salient object detection (SOD) methods. However, RGB@$\mathrm{D}$pairs may not always be readily accessible. Therefore, in this paper, we introduce a SOD framework comprising two modules: the Depth Information Generation (DIG) module and the RGB-Generated Depth fusion (RGB-GD fusion) module. The proposed DIG module is designed using Generative Adversarial Network (GAN) to generate depth maps from input RGB images. The resulting generated depth map is then integrated with RGB data using RGB-GD fusion module to enhance SOD. The proposed DIG module comprises a generator and a discriminator, which learn the mapping from RGB to depth. To enhance depth generation accuracy further, a depth enhancer is introduced in the DIG module. Both qualitative and quantitative evaluations validate the effectiveness of the DIG module compared to existing state-of-the-art depth generators. Moreover, we validate the utility of the generated depth maps in the SOD task through qualitative and quantitative comparisons with state-of-the-art RGB-D SOD methods on RGB-D benchmark datasets. Experimental results demonstrate the applicability of the proposed method when depth maps are not readily available.