Abstract:The ability to present three-dimensional (3D) scenes with continuous depth sensation has a profound impact on virtual and augmented reality, human–computer interaction, education and training. Computer-generated holography (CGH) enables high-spatio-angular-resolution 3D projection via numerical simulation of diffraction and interference<a href="/articles/s41586-020-03152-0#ref-CR1">1</a>. Yet, existing physically based methods fail to produce holograms with both per-pixel focal control and accurate occlusion<a href="/articles/s41586-020-03152-0#ref-CR2">2</a>,<a href="/articles/s41586-020-03152-0#ref-CR3">3</a>. The computationally taxing Fresnel diffraction simulation further places an explicit trade-off between image quality and runtime, making dynamic holography impractical<a href="/articles/s41586-020-03152-0#ref-CR4">4</a>. Here we demonstrate a deep-learning-based CGH pipeline capable of synthesizing a photorealistic colour 3D hologram from a single RGB-depth image in real time. Our convolutional neural network (CNN) is extremely memory efficient (below 620 kilobytes) and runs at 60 hertz for a resolution of 1,920 × 1,080 pixels on a single consumer-grade graphics processing unit. Leveraging low-power on-device artificial intelligence acceleration chips, our CNN also runs interactively on mobile (iPhone 11 Pro at 1.1 hertz) and edge (Google Edge TPU at 2.0 hertz) devices, promising real-time performance in future-generation virtual and augmented-reality mobile headsets. We enable this pipeline by introducing a large-scale CGH dataset (MIT-CGH-4K) with 4,000 pairs of RGB-depth images and corresponding 3D holograms. Our CNN is trained with differentiable wave-based loss functions<a href="/articles/s41586-020-03152-0#ref-CR5">5</a> and physically approximates Fresnel diffraction. With an anti-aliasing phase-only encoding method, we experimentally demonstrate speckle-free, natural-looking, high-resolution 3D holograms. Our learning-based approach and the Fresnel hologram dataset will help to unlock the full potential of holography and enable applications in metasurface design<a href="/articles/s41586-020-03152-0#ref-CR6">6</a>,<a href="/articles/s41586-020-03152-0#ref-CR7">7</a>, optical and acoustic tweezer-based microscopic manipulation<a href="#ref-CR8">8</a>,<a href="#ref-CR9">9</a>,<a href="/articles/s41586-020-03152-0#ref-CR10">10</a>, holographic microscopy<a href="/articles/s41586-020-03152-0#ref-CR11">11</a> and single-exposure volumetric 3D printing<a href="/articles/s41586-020-03152-0#ref-CR12">12</a>,<a href="/articles/s41586-020-03152-0#ref-CR13">13</a>.

Large-Field Holographic Projection System Based On Deep-Learning Acceleration Calculation

Multiplane Digital Holography Based on Extrapolation Iterations

Large Field-of-view Lensless Holographic Dynamic Projection System with Uniform Illumination and U-net Acceleration

Holographic near-eye display system based on double-convergence light Gerchberg-Saxton algorithm.

Computer-generated holographic near-eye display system based on LCoS phase only modulator

Deep-learning-aided multi-focal hologram generation

360-degree large-scale multiprojection light-field 3D display system.

Lensless holographic dynamic projection system based on weakly supervised learning

Generating Real-Scene Hologram Through Light Field Imaging and Deep Learning

Fast Hologram Generation for Active Holographic Display

High-quality and high-speed computer-generated holography via deep-learning-assisted bidirectional error diffusion method

High-Quality Holographic 3D Display System Based on Virtual Splicing of Spatial Light Modulator

Generating Multi‐Depth 3D Holograms Using a Fully Convolutional Neural Network

Large-Sized Light Field Three-Dimensional Display Using Multi-Projectors And Directional Diffuser

Extended Depth-of-field in Holographic Imaging Using Deep-Learning-based Autofocusing and Phase Recovery

Unsupervised deep neural network for fast computer-generated holography with continuous depth control

Dynamic three-dimensional light-field display with large viewing angle based on compound lenticular lens array and multi-projectors

A Deep Learning Approach for the Fast Generation of Acoustic Holograms.

High-fidelity, Model-Driven Deep Learning Network for Phase-Only Computer-Generated Holography (conference Presentation)

Towards real-time photorealistic 3D holography with deep neural networks

DGE-CNN: 2D-to-3d Holographic Display Based on a Depth Gradient Extracting Module and ZCNN Network.