Weakly-Supervised Salient Object Detection With Saliency Bounding Boxes.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Salient Object Detection (SOD) is crucial for image understanding.
• Existing weakly-supervised SOD methods often struggle with precise localization.
• Saliency bounding boxes offer a concise representation of salient object locations.

Purpose of the Study:

• To propose a novel weakly-supervised salient object detection method using saliency bounding boxes.
• To generate pixel-level pseudo ground truth saliency maps from bounding box annotations.
• To improve the performance of SOD models with limited supervision.

Main Methods:

• Utilizing unsupervised SOD methods to generate initial saliency maps.
• Refining initial saliency maps to address over/under-prediction issues.
• Employing a multi-task map refinement network trained with saliency bounding boxes.
• Using generated pseudo saliency maps for training the final SOD detector.

Main Results:

• The proposed method successfully generates pixel-level pseudo ground truth saliency maps.
• Iterative refinement using a multi-task network enhances map accuracy.
• The final SOD detector trained with pseudo maps achieves superior performance.
• Experimental results demonstrate outperformance against state-of-the-art weakly-supervised SOD methods.

Conclusions:

• Weak supervision for SOD using saliency bounding boxes is effective.
• The proposed pseudo ground truth generation and refinement strategy significantly boosts SOD accuracy.
• This approach offers a viable alternative for training SOD models when pixel-level annotations are unavailable.