Structured light scanners triangulate a projected pattern against stereo or single-camera views, and the 2026 commercial and open-source landscape spans desktop units under 0.01 m^3 working volume to 2 m^3 robotic cells resolving 0.05-0.1 mm point spacing [S2][S3].
Buyers should treat the choice as a four-axis decision: stand-off distance, baseline (projector-to-camera), pattern-encoding family, and host compute path. Open-source benchmarks such as [S1] slBenchmark and 2026-vintage reconstruction codebases including 3DUNDERWORLD-SLS-GPU_CPU [S2] and Triangulum3D [S3] frame the same decision axes that industrial OEMs use, and the [3DUNDERWORLD] pipeline is built around OpenCV2 plus CUDA 6.0+ with a CMake `ENABLE_CUDA=off` fallback for CPU-only hosts [S2].
Working volume and stand-off: the first filter
Working volume for a structured light scanner is defined by the intersection of projector field-of-view and camera field-of-view, and shrinks geometrically with the baseline ratio; a 0.3 m stand-off with a 0.15 m baseline typically yields a 200 x 150 x 150 mm capture zone at acceptable depth-of-field [S1].
For desktop metrology (jewellery, electronics, dental), 100-300 mm stand-off with 0.1-0.2 mm stated accuracy is the dominant 2026 SKU class. Mid-range body-scanning and reverse-engineering cells run 0.8-1.5 m stand-off, while robotic cells for automotive sheet metal reach 1.5-2.5 m. Used industrial units from older GOM and Steinbichler lines surfaced in 2024-2026 secondary markets at the US$1,400-3,000 band [S7], but firmware support for current CAD packages is the gating risk on those assets. The compact 3DUNDERWORLD code path targets the same single-shot-to-point-cloud workflow that 3D scanner selection criteria lays out for buyers.
Pattern encoding: binary Gray, multi-frequency, and sinusoidal fringe
Binary Gray-code projection remains the most common 2026 encoding for sub-0.1 mm metrology because it gives an absolute per-pixel code with log2(N) patterns for N stripes, and Triangulum3D [S3] implements exactly that scheme with explicit binary-coded stripe sequences. Multi-frequency (heterodyne) and sinusoidal phase-shifting encodings trade pattern count for sub-pixel phase interpolation, and they are the default in modern fringe-projection industrial heads where 0.01-0.05 mm point spacing is required.
Compared on three criteria, the three families line up as follows: (1) pattern count per scan, where binary Gray needs 8-12 frames for a 1024-period code, multi-frequency needs 12-18, and sinusoidal phase-shifting needs 3-5 with N-step averaging; (2) robustness to ambient light and surface albedo, where sinusoidal phase-shifting is the strongest because phase unwrapping tolerates low SNR better than edge decoding; (3) per-pixel absolute code, where binary Gray wins outright, which matters for discontinuous surfaces such as turbine blades and stamped sheet. Buyers mapping a 2026 selection should default to sinusoidal phase-shifting for smooth, diffuse, mid-volume targets, and to binary or multi-frequency when the part has holes, undercuts, or mixed reflectance. The pattern-encoding axis is treated in more depth in Structured Light Scanner Selection: Working Volume, Baseline, and Pattern Encoding, which complements the 3DUNDERWORLD reference design [S2].
Baseline, triangulation angle, and depth resolution
Depth resolution in a structured light scanner scales linearly with baseline length and inversely with the square of stand-off; doubling the projector-to-camera baseline roughly halves the depth uncertainty at a fixed pixel resolution [S1].
Industrial heads in 2026 use baselines of 0.15-0.4 m at 0.3-0.5 m stand-off for 0.02-0.05 mm volumetric accuracy. Open rigs such as neuvision [S4] expose baseline and convergence as calibration parameters, which is exactly the discipline that slBenchmark [S1] codifies for repeatable research comparisons. Buyers should reject any spec sheet that lists accuracy without the matching baseline, stand-off, and working-volume triple. The `depth-of-view` correction in slBenchmark (commit message "Divide the depth of view by 2 to get the right number", master branch) is a small but telling artefact: published numbers often hide a factor-of-2 disagreement between manufacturers and that the buyer must reconcile in writing [S1].
Compute path: CPU, GPU, and embedded SoC
Reconstruction time for a single 1.2 MP scan frame pair is typically 0.3-2 s on a current desktop CPU and 30-150 ms on a CUDA-class GPU, with the 3DUNDERWORLD codebase demonstrating the same speedup pattern via its dual `ImageProcessor` and `Reconstructor` blocks [S2].
Embedded hosts (Jetson Orin, industrial PC with TPU) are sufficient when the projected pattern count is held to 6-8 and resolution stays at 1 MP, which is the typical embedded-3D-scanner niche. For shop-floor cells where line-scan cameras and 4K projectors are in play, x86 + discrete GPU is still the default in 2026 because the phase-unwrapping and correspondence-search kernels scale poorly on low-memory embedded parts. Buyers evaluating open pipelines should confirm OpenCV2 compatibility and the presence of a `cmake .. -DENABLE_CUDA=off` fallback, because that single flag determines whether a buyer without a workstation GPU can still run the pipeline [S2].
Calibration, open-source reference designs, and reproducibility
Reproducibility is the single biggest 2026 differentiator in open structured light work: slBenchmark [S1] is built around the explicit goal that "adopters of the framework will be able to reliably propose, compare, reproduce and verify different structured light approaches", and it uses camera and projector calibration as a first-class input.
Three open-source reference designs are worth tracking in 2026: 3DUNDERWORLD-SLS-GPU_CPU (147 commits, CPU+GPU pipeline, OpenCV2 + CUDA 6.0+, Doxygen docs, Google Test + lcov coverage) [S2]; Triangulum3D (33 commits, binary-coded stripes, 12 forks, 17 stars) [S3]; and neuvision (259 commits, structured-light scanner with apps, lib, and 3rdparty submodules) [S4]. The smaller `structured-light-scanner` repo [S5] (9 commits, single `main.py` with `structuredlight_reconstruction.py` and `utils.py`) is a teaching reference rather than a metrology platform. For pure research benchmarking, slBenchmark's 16-commit history [S1] documents the experimental-design discipline that any in-house scanner team should adopt before signing a purchase order. The ecosystem is the same one surveyed in 3D Scanner Selection Criteria: A Spec-First Decision Map for 2026, which gives a broader view of how structured light fits against laser-line, ToF, and photogrammetry alternatives.
Who should NOT pick a structured light scanner in 2026
Structured light is the wrong tool when the target is dark, specular, or moving faster than 1 m/s, because projection-based encoding depends on a static, diffusely reflective surface during the full pattern sequence. [S2]
Outdoor use under direct sunlight is also a poor fit: ambient IR/visible wash typically exceeds the projector's modulation depth by 10x or more, and phase unwrapping breaks down above roughly 10 klx. For those cases, laser-line triangulation or time-of-flight remains the right call. Buyers also need to reject any unit whose accuracy claim is not paired with a stated working volume, a stated baseline, and a stated stand-off, because the three are mathematically coupled and a single number is meaningless. Trackable 2026 signals: GitHub activity on slBenchmark, neuvision, and 3DUNDERWORLD-SLS-GPU_CPU (commit cadence over the next two quarters) and the next release of vendor firmware supporting the new Gocator and Hexagon IFS lines referenced in the secondary market [S7].
Spec-level background on the components involved: structured light scanner, linear guide, and crossed roller guide.