Voxel Grid-Based Depth Recovery From Monocular Structured Light.

Abstract

We introduce an approach for depth estimation from images captured by monocular structured light systems, building upon the matching-free differentiable rendering paradigm with a key architectural contribution: replacing the implicit signed distance function (SDF) geometry representation with an explicit density voxel grid. Unlike methods that rely on image matching, our technique leverages the known pattern projection and camera setup to guide geometry optimization through differentiable volume rendering, without any explicit correspondence search. We isolate and optimize the geometry field by treating projected patterns as a known color field during rendering, thereby enabling faster convergence and higher-quality results. We further introduce a set of complementary losses-a patch-based SSIM loss, a distortion loss for unimodal density, and a surface-aware color loss-that are specifically designed for the voxel grid representation in monocular structured light settings. Experiments show that our method outperforms matching-based techniques in few-shot scenarios, reducing average depth error by over 30% on both synthetic and real-world scenes. Compared to prior matching-free methods using SDFs, our voxel grid-based framework avoids smoothness-induced artifacts at sharp edges, captures fine geometric discontinuities more faithfully, and trains nearly seven times faster. We provide a systematic analysis of the trade-offs between explicit voxel grids and implicit SDFs, clarifying why the explicit representation is better suited to monocular structured light depth recovery.