Weakly Supervised RGB-D Salient Object Detection With Prediction Consistency Training and Active Scribble Boosting.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• RGB-D salient object detection (SOD) offers improved performance in complex scenes over RGB-only methods.
• Current state-of-the-art SOD models require extensive pixel-wise annotations, which are costly and time-consuming.
• Weakly supervised learning presents a viable alternative to reduce annotation burden.

Purpose of the Study:

• To investigate weakly supervised RGB-D SOD using annotator-friendly scribble annotations.
• To develop a novel network architecture that preserves object structure information from sparse annotations.
• To enhance SOD performance by leveraging complementary edge information and active learning strategies.

Main Methods:

• Utilized scribble annotations as supervision signals for training RGB-D SOD models.
• Incorporated dual-modal edge guidance, a dual-edge detection module, and a modality-aware feature fusion module.
• Implemented a prediction consistency training scheme and an active scribble boosting strategy for improved performance.

Main Results:

• The proposed method demonstrates superior performance on seven benchmark datasets.
• Achieved competitive results compared to fully supervised state-of-the-art methods using only scribble annotations.
• Validated the effectiveness of dual-modal edge guidance and active scribble boosting in enhancing SOD accuracy.

Conclusions:

• Weakly supervised RGB-D SOD using scribbles is a cost-effective and efficient alternative to fully supervised methods.
• The proposed network architecture effectively preserves structural information and fuses multi-modal features.
• Active scribble boosting significantly improves SOD performance with minimal annotation effort.