A Large-Depth-Range Layer-Based Hologram Dataset for Machine Learning-Based 3D Computer-Generated Holography

Technical Summary: A Large-Depth-Range Layer-Based Hologram Dataset for Machine Learning-Based 3D Computer-Generated Holography

Problem Statement
Machine learning-based computer-generated holography (ML-CGH) has shown rapid progress in enabling real-time 3D displays for augmented and virtual reality. However, the development of robust, generalizable ML-CGH models is currently hindered by a scarcity of high-quality, large-scale training datasets. Existing public datasets, such as MIT-CGH-4K, suffer from significant limitations: they offer narrow depth ranges (e.g., 6 mm), specific optical configurations optimized for particular hardware, and limited resolution diversity. Furthermore, many private datasets used in recent studies are not publicly available, preventing reproducibility and fair benchmarking. There is a critical gap in the availability of a dataset that simultaneously offers high spatial resolution and a large depth range suitable for training next-generation ML-CGH systems.

Methodology
To address these limitations, the authors introduce KOREATECH-CGH, a publicly available dataset comprising 6,000 pairs of RGB-D images and complex holograms. The dataset generation pipeline involves the following key components:

1. Scene Construction: Scenes are synthesized using 100 objects from the Google Scanned Objects (GSO) dataset. Objects are randomly scaled, rotated, and positioned within a camera field of view. The depth distribution spans a range up to the theoretical limits of the angular spectrum method (ASM) for the given resolution, reaching up to 80 mm.
2. Rendering: RGB images and depth maps are rendered using OptiX 8.0 with orthographic projection.
3. Hologram Generation (AP-LBM): The core innovation is a novel Amplitude-Projection-Enhanced Layer-Based Method (AP-LBM). This method improves upon existing silhouette-masking layer-based approaches (SM-LBM) through two mechanisms:
   • Advanced Silhouette Masking: It employs bidirectional masking to filter wavefronts more effectively, reducing contour line aliasing and shadow artifacts between focal planes.
   • Amplitude Projection: To mitigate defocus blur in large-depth-range scenes, the method projects the image content onto the amplitude component of the hologram wavefield at each corresponding depth layer while preserving the phase. This iterative process refines the hologram by replacing amplitude components at specific depths with the target image content, significantly enhancing reconstruction fidelity in both near and far focal planes.
4. Dataset Specifications: The dataset covers four resolutions (256×256 to 2048×2048) with a fixed pixel pitch of 3.6 µm and wavelengths corresponding to RGB channels (638 nm, 532 nm, 450 nm).

Key Contributions

• KOREATECH-CGH Dataset: A public dataset of 6,000 high-resolution, large-depth-range RGB-D and hologram pairs, designed to facilitate the training and evaluation of ML-CGH models.
• AP-LBM Algorithm: A novel generation technique that integrates amplitude projection with advanced layer-based masking to achieve superior reconstruction quality over large depth ranges compared to traditional methods.
• Evaluation Framework: The introduction of Focal Image Projection (FIP) as a quantitative metric. Unlike standard PSNR/SSIM comparisons that may penalize natural defocus blur, FIP aggregates amplitude information from the focal stack using depth-aware masking to compare only in-focus regions against the rendered target, providing a more accurate measure of holographic fidelity.

Results

• Hologram Quality: Quantitative evaluations using FIP-based PSNR and SSIM demonstrate that AP-LBM outperforms both the original SM-LBM and the advanced bidirectional masking method (ADV-LBM). AP-LBM achieved a maximum PSNR of 27.01 dB and an SSIM of 0.87, surpassing the optimized silhouette-masking method by 2.03 dB and 0.04 SSIM, respectively.
• Depth Range Analysis: While image quality metrics naturally degrade as the depth range increases (e.g., dropping from ~30 dB at 1 mm to ~24 dB at 20 mm), AP-LBM maintains usable quality even at the maximum tested depth of ~20.3 mm, significantly outperforming previous methods in this regime.
• Layer Count Sensitivity: Experiments varying the number of layers (from 10 to 10,000) indicate that increasing layers improves overall quality, with diminishing returns beyond 100 layers for minimum scores, though 10,000 layers yielded the highest average and maximum metrics.
• Machine Learning Applications: The dataset was used to train several models from scratch, including TensorHolography, U-Net, Swin-Unet, and super-resolution models (H2HSR). All models demonstrated reasonable performance, confirming the dataset's utility for training both hologram generation and upscaling tasks. Notably, the authors observed that model performance varied based on the dataset's depth configuration, suggesting that the larger depth range of KOREATECH-CGH presents a more challenging and realistic training environment than previous datasets.

Significance
The paper claims that KOREATECH-CGH fills a critical void in the ML-CGH community by providing a reproducible, high-quality foundation for research. By offering a dataset with a significantly larger depth range and higher resolutions than existing public options, it enables the development of more robust and generalizable ML-CGH systems. The authors posit that this resource will catalyze progress in next-generation 3D displays, particularly for applications requiring immersive, depth-aware visualization such as AR and VR, by allowing researchers to benchmark and train models on data that better reflects the complexities of real-world 3D scenes.