Synthesizing Supervision for Learning Deep Saliency Network without Human Annotation.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks have advanced salient object detection (SOD).
• Current SOD models require expensive pixel-level annotations for training.
• There is a need for annotation-free methods to reduce costs and time.

Purpose of the Study:

• To develop an early framework for training deep salient object detectors without human annotation.
• To explore novel supervision synthesis schemes for annotation-free learning.
• To reduce the reliance on pixel-level ground truth masks.

Main Methods:

• A novel supervision synthesis scheme using "knowledge source transition" and "supervision by fusion".
• Dynamic exploration of external and internal knowledge sources for supervision cues.
• A two-stream fusion mechanism for implementing the supervision synthesis process.

Main Results:

• The annotation-free deep salient object detector achieved performance comparable to fully supervised methods (within a 3% gap).
• The framework successfully trained models without requiring human-annotated masks.
• The learned detector assisted weakly supervised semantic segmentation, reducing supplementary supervision needs.

Conclusions:

• Annotation-free learning is feasible for deep salient object detection.
• The proposed framework offers a cost-effective alternative to traditional supervised methods.
• This approach has implications for related tasks like weakly supervised semantic segmentation.