A Thorough Benchmark and a New Model for Light Field Saliency Detection.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Existing light field saliency detection datasets lack sufficient data, diversity, and detailed annotations.
• This hinders progress in the field of light field saliency detection.

Purpose of the Study:

• To introduce a large-scale, diverse light field dataset (PKU-LF) with comprehensive annotations.
• To propose a novel deep learning architecture for enhanced light field saliency detection.

Main Methods:

• Construction of the PKU-LF dataset with 5,000 light fields, diverse scenes, and multiple annotation types (scribbles, bounding boxes, object/instance-level, edges).
• Systematic evaluation and benchmarking of 16 existing 2D, 3D, and 4D saliency detection methods.
• Development of a novel Symmetric Two-Stream Architecture (STSA) network incorporating Focalness Interweavement Module (FIM) and Partial Decoder Modules (PDM).

Main Results:

• The PKU-LF dataset provides a unified platform for algorithm comparison and enables investigation of new attention modeling tasks.
• The proposed STSA network significantly outperforms existing methods in light field saliency detection.
• Comprehensive analysis establishes a thorough benchmark for light field saliency detection methods.

Conclusions:

• The PKU-LF dataset addresses critical limitations in existing resources, fostering advancements in light field saliency detection.
• The STSA network offers a promising direction for accurate and efficient light field saliency prediction.