Pushing the Boundaries of Salient Object Detection: A Denoising-Driven Approach

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient Object Detection (SOD) identifies prominent image regions, crucial for various vision tasks.
• Existing discriminative SOD methods falter in complex scenes with ambiguous object-background distinctions.
• Current RGB-D fusion methods often struggle with direct merging of appearance and depth data.

Purpose of the Study:

• To develop a diffusion-based model (DiffSOD) for robust salient object detection in RGB and RGB-D images.
• To improve SOD performance in challenging scenarios with low contrast and complex textures.
• To establish a new benchmark for diffusion-based dense prediction models in visual saliency.

Main Methods:

• Employs a diffusion framework utilizing a noise-to-image denoising process.
• Treats RGB (appearance) and depth (structure) as distinct conditional inputs to a UNet architecture.
• Introduces specialized appearance and structure control adapters, plus a quality-aware filter for depth data.

Main Results:

• DiffSOD significantly outperforms existing RGB and RGB-D saliency detection methods on benchmark datasets.
• Achieved average performance improvements of 1.5% for RGB and 1.2% for RGB-D data.
• Demonstrates enhanced saliency detection capabilities, particularly in complex visual scenes.

Conclusions:

• DiffSOD offers a superior approach to salient object detection by leveraging diffusion models.
• The conditional guidance and quality-aware filtering effectively handle complex scenes and varying depth data quality.
• Presents a new state-of-the-art for diffusion-based dense prediction in visual saliency detection.