A New Parallel Intelligence Based Light Field Dataset for Depth Refinement and Scene Flow Estimation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Light field images offer superior structural information for computer vision tasks compared to 2D images.
• Acquiring real-world light field data is challenging due to instrument calibration issues.
• Existing light field datasets are often small, limiting their use in advanced methods like transformers, and dynamic scene studies are scarce.

Purpose of the Study:

• Introduce a novel, large-scale static light field dataset with comprehensive ground truth for tasks like disparity and depth estimation.
• Release a dynamic light field dataset with dense 3D motion ground truth for precise motion estimation and object tracking.
• Evaluate the dataset's utility for disparity estimation and angular super-resolution.

Main Methods:

• Developed a static light field dataset with 50 scenes, 8-10 perspectives each, and ground truth for disparities, depths, normals, segmentations, and poses.
• Created a dynamic light field dataset with 150 frames per scene, providing dense 3D motion ground truth for optical flow.
• Utilized DistgDisp and DistgASR methods to decouple angular and spatial domains for performance evaluation.

Main Results:

• The new static dataset is larger than current mainstream datasets for depth estimation refinement.
• The dynamic dataset provides dense 3D motion ground truth for pixel-level motion estimation.
• Experimental results demonstrate the dataset's effectiveness in evaluating disparity estimation and angular super-resolution.

Conclusions:

• The proposed light field datasets significantly advance research in static and dynamic computer vision tasks.
• The datasets enable more accurate motion estimation, disparity refinement, and angular super-resolution.
• This work provides a valuable resource for developing and benchmarking advanced light field analysis algorithms.