Exploring better sparsely annotated shadow detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Sparsely annotated image segmentation offers reduced labeling costs but faces challenges in weakly-supervised shadow detection.
• Existing methods struggle with weak supervision diffusion and poor structure recovery, leading to a performance gap.

Purpose of the Study:

• To propose a one-stage weakly-supervised learning framework for improved sparsely annotated shadow detection.
• To alleviate challenges of weak supervision diffusion and poor structure recovery in shadow detection.

Main Methods:

• Developed a semantic affinity module (SAM) for adaptive propagation of scribble supervision using gradient diffusion.
• Introduced a feature-guided edge-aware loss to enhance shadow boundary perception.
• Implemented an intensity-guided structure consistency loss for improved model generalization.

Main Results:

• The proposed framework significantly outperforms previous weakly-supervised shadow detection methods.
• Achieved competitive performance compared to state-of-the-art fully-supervised methods on benchmark datasets.

Conclusions:

• The novel framework effectively addresses key challenges in sparsely annotated shadow detection.
• Demonstrated superior performance and generalization ability, narrowing the gap with fully-supervised approaches.