Deep Hypersphere Feature Regularization for Weakly Supervised RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Salient object detection typically requires dense pixel-level annotations, which are labor-intensive.
• Existing weakly supervised methods often focus on output-space supervision, limiting their effectiveness.

Purpose of the Study:

• To develop a weakly supervised salient object detection method using RGB-D data with minimal annotation effort.
• To improve the discrimination between salient and non-salient objects by regularizing the latent space.

Main Methods:

• Utilizes scribble-based labels for weak supervision, significantly reducing annotation costs.
• Employs latent space regularization to enhance feature discrimination.
• Introduces a contour detection branch for precise object boundary refinement.
• Incorporates a Cross-Padding Attention Block (CPAB) to capture long-range feature dependencies.

Main Results:

• Outperforms existing weakly supervised salient object detection methods.
• Achieves performance on par with several state-of-the-art fully supervised models.
• Demonstrates effectiveness across seven benchmark datasets.

Conclusions:

• The proposed weakly supervised approach offers a practical and efficient solution for salient object detection.
• Latent space regularization and contour constraints contribute to high-accuracy salient object detection.
• The method provides a competitive alternative to fully supervised techniques, especially when annotations are limited.