[Paper Review] A Study of Efficient Light Field Subsampling and Reconstruction Strategies

Yang Chen, Martin Alain+1 more|arXiv (Cornell University)|Aug 11, 2020

Ocular and Laser Science ResearchMedicine20 citations

TL;DR

This paper investigates efficient light field subsampling and reconstruction strategies to improve angular resolution for practical applications. It evaluates row-wise, column-wise, and checkerboard subsampling patterns, and compares six reconstruction methods for generating intermediate views from sparse corner views. The key finding is that row-wise subsampling followed by column-wise reconstruction yields the best performance, achieving PSNR of 37.34 dB and SSIM of 0.9886 at full angular density.

ABSTRACT

Limited angular resolution is one of the main obstacles for practical applications of light fields. Although numerous approaches have been proposed to enhance angular resolution, view selection strategies have not been well explored in this area. In this paper, we study subsampling and reconstruction strategies for light fields. First, different subsampling strategies are studied with a fixed sampling ratio, such as row-wise sampling, column-wise sampling, or their combinations. Second, several strategies are explored to reconstruct intermediate views from four regularly sampled input views. The influence of the angular density of the input is also evaluated. We evaluate these strategies on both real-world and synthetic datasets, and optimal selection strategies are devised from our results. These can be applied in future light field research such as compression, angular super-resolution, and design of camera systems.

Motivation & Objective

• To identify optimal light field subsampling and reconstruction strategies for improved angular resolution in practical applications.
• To evaluate the impact of different subsampling patterns—row-wise, column-wise, and checkerboard—on reconstruction quality at a fixed sampling ratio.
• To compare multiple reconstruction strategies for generating intermediate views from four corner views in a 3×3 grid.
• To analyze the effect of angular density on reconstruction quality using a multi-stage reconstruction approach.
• To provide actionable insights for light field compression, camera system design, and super-resolution research.

Proposed method

• A state-of-the-art view synthesis method (SepConv) was selected as the benchmark for evaluating all subsampling and reconstruction strategies.
• Subsampling strategies were tested on real-world (HCI, Lytro) and synthetic (Stanford) light field datasets, with views cropped to 512×512 resolution for computational efficiency.
• Reconstruction strategies included 2D H-V (horizontal then vertical), 2D V-H (vertical then horizontal), diagonal methods, and 4D omni/diagonal approaches using a retrained SepConv model.
• Multi-stage reconstruction was applied recursively to generate dense light fields from sparse inputs at three angular density levels (Level 1–3), with results averaged across all views.
• Performance was quantitatively evaluated using PSNR and SSIM on RGB images, with mean scores computed across all views of each light field.
• The benchmark model was fine-tuned for 2D strategies and retrained for 4D strategies to ensure fair comparison across methods.

Experimental results

Research questions

• RQ1 Which subsampling pattern—row-wise, column-wise, or checkerboard—produces the best reconstruction quality at a fixed sampling ratio?
• RQ2 Which reconstruction strategy—2D H-V, 2D V-H, diagonal, or 4D methods—yields the highest quality intermediate views from four corner views?
• RQ3 How does angular density affect the quality of reconstructed light fields when using a multi-stage reconstruction pipeline?
• RQ4 Can the optimal subsampling and reconstruction strategy be generalized across diverse light field datasets with varying capture geometries?
• RQ5 What are the trade-offs between subsampling efficiency and reconstruction fidelity in light field applications?

Key findings

• Row-wise subsampling consistently outperformed column-wise and checkerboard patterns, achieving a mean PSNR of 39.74 dB and SSIM of 0.9925 across datasets.
• The 2D H-V reconstruction strategy (row-wise then column-wise interpolation) achieved the highest PSNR (37.42 dB) and SSIM (0.9884), outperforming other 2D and 4D strategies.
• Checkerboard subsampling performed worst, likely due to suboptimal model adaptation when retraining SepConv for 4-view inputs.
• As angular density decreased from Level 1 to Level 3, PSNR dropped from 37.34 dB to 31.79 dB, and SSIM from 0.9886 to 0.9613, confirming a clear trade-off between sparsity and quality.
• Visual analysis revealed occlusion artifacts in diagonal reconstruction strategies, particularly around depth discontinuities such as the sword tip.
• The multi-stage 2D H-V strategy was found to be optimal for reconstructing dense light fields from sparse corner views, especially when input views are spatially well-distributed.