High-speed 3D light field sensing via snapshot compressive acquisition and domain-adaptive deep equilibrium reconstruction

Constrained by the finite space-bandwidth product of sensors, 3D light field (LF) sensing using consumer LF cameras is often restricted to single-digit frame rates. To bridge this gap, we present a physical snapshot compressive acquisition system that encodes multiple 5D LF frames into a single 4D measurement via a digital micromirror device (DMD). To recover high-fidelity LF data, we propose the domain-adaptive multi-stage deep equilibrium (DAM-DEQ) framework. This model replaces uniform processing strategies with a structure-aware design that first integrates spatial and angular priors, followed by a multi-domain fusion stage where parallel spatial, angular, and epipolar features are weighted adaptively and refined via spatiotemporal convolution. Experiments on synthetic (Sintel), real (Lytro), and custom array-camera datasets demonstrate that our method outperforms existing techniques and DEQ baselines in terms of PSNR and SSIM. Furthermore, our system achieves effective reconstruction fidelity even on physical measurements degraded by optical aberrations and mask misalignment, confirming the practical viability of the proposed hardware-algorithm co-design.