Robust Perception and Precise Segmentation for Scribble-Supervised RGB-D Saliency Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Pixel-wise annotations for salient object detection (SOD) are labor-intensive.
• Weakly supervised methods using scribbles reduce annotation burden but face performance drops.
• Existing scribble-based SOD methods struggle with insufficient training data richness and poor object structure.

Purpose of the Study:

• To propose a novel scribble-based weakly supervised RGB-D salient object detection (SOD) method.
• To address the challenges of weak richness of pixel training samples (WRPS) and poor structural integrity of salient objects (PSIO) in scribble-based SOD.
• To achieve robust feature learning and precise object segmentation with reduced annotation effort.

Main Methods:

• A dynamic searching process module for multi-scale and multi-modal feature fusion in RGB-D SOD.
• A dual-branch consistency learning mechanism to generate sufficient pixel training samples.
• An edge-region structure-refinement loss to recover object structural information for precise segmentation.

Main Results:

• The proposed method effectively alleviates WRPS through robust feature learning and pseudo-label generation.
• The edge-region structure-refinement loss successfully addresses PSIO, improving segmentation accuracy.
• Experimental results on eight datasets demonstrate superior performance compared to other scribble-based SOD methods.

Conclusions:

• The developed scribble-based weakly supervised RGB-D SOD method significantly reduces annotation burden while maintaining high performance.
• The method achieves comparable results to fully supervised state-of-the-art SOD approaches.
• The proposed techniques offer a promising direction for efficient and effective salient object detection.