Learning to Detect Salient Object With Multi-Source Weak Supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pixel-level annotations for saliency detection are costly.
• Weak supervision offers a cost-effective alternative but often lacks sufficient information.
• Existing weakly supervised methods struggle to achieve high performance.

Purpose of the Study:

• To develop a unified two-stage framework for training saliency detection models using diverse weak supervision sources.
• To overcome the limitations of single weak supervision by integrating multiple data types.
• To improve the performance of saliency detection models in a cost-efficient manner.

Main Methods:

• A two-stage framework combining classification (CNet) and caption generation (PNet) networks.
• Utilizing attention transfer and coherence losses to guide feature learning.
• Creating complementary training datasets (natural images with noisy labels, synthesized images) for the final saliency prediction network (SNet).

Main Results:

• The proposed method successfully trains saliency detection models using category labels, captions, web images, and unlabeled images.
• The two-stage framework effectively transfers knowledge from auxiliary tasks to saliency prediction.
• Experimental results demonstrate competitive performance against unsupervised, weakly supervised, and some supervised methods.

Conclusions:

• The unified two-stage framework offers a powerful approach for weakly supervised saliency detection.
• Integrating multiple weak supervision sources enhances model robustness and accuracy.
• This method provides a viable and effective alternative to costly pixel-level annotations.