Weakly-Supervised Salient Object Detection on Light Fields.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing salient object detection (SOD) methods primarily use RGB images, neglecting rich light field data.
• Current light field SOD approaches often require extensive pixel-level manual annotations.
• This limits the practical application of advanced SOD techniques.

Purpose of the Study:

• To develop a novel weakly-supervised learning framework for salient object detection in light field images.
• To overcome the limitations of pixel-level annotations by utilizing bounding box information.
• To leverage the multi-view and depth information inherent in light fields for improved saliency detection.

Main Methods:

• Proposed a ground truth label hallucination method to generate pixel-level pseudo saliency maps from bounding box annotations.
• Introduced a fusion attention module to effectively process multi-view light field data, calibrating spatial and channel-wise representations.
• Developed a two-branch network (RGB and Focal) for weakly-supervised salient object detection.

Main Results:

• The ground truth hallucination method successfully generated high-quality pseudo saliency maps.
• The fusion attention module effectively focused on informative features and reduced redundancy from multi-view inputs.
• The proposed weakly-supervised method achieved superior performance compared to existing weakly-supervised and most fully supervised methods.

Conclusions:

• Weakly-supervised learning with bounding box annotations is a viable and efficient approach for salient object detection in light field images.
• The proposed framework effectively utilizes light field data, including depth and multi-view information, for accurate saliency detection.
• This method offers a practical solution for salient object detection, reducing annotation costs and improving performance on complex scenes.