Home Reviews RealityCapture Review 2018: Photogrammetry Workflow and Results

March 15, 2018

RealityCapture Review 2018: Photogrammetry Workflow and Results

RealityCapture review 2018 explained: photogrammetry workflow, image alignment, texture export, and what the old results do and don’t prove.

Summary

This RealityCapture review 2018 article is best read as an archival explainer, not a current buying guide. The underlying 3D Mag review was originally published on March 15, 2018, the page later showed an update date of July 28, 2026, and the displayed result models were made in June 2017 and were not reprocessed for the later refresh. [1]

That makes the page useful as evidence of one historical RealityCapture photogrammetry workflow and one reviewer’s outputs, but not as a controlled benchmark. The review does not publish a metrology-grade comparison protocol or a universal accuracy number, so the safest reading remains: no reliable figure found for a single all-purpose performance claim. [1]

RealityCapture review 2018 — what the original test actually is

The key fact about a RealityCapture review 2018 page is that it combines three time layers: older result models, a March 2018 publication date, and a much later page refresh. If those layers are not separated, a historical review can easily be mistaken for a current retest. The 3D Mag page is most useful as an archival record of what the reviewer saw at the time, on that machine, with those datasets and that presentation context. [1]

Date	What it refers to	Reader takeaway
June 2017	Result models used in the article	Historical outputs, not reprocessed for the later page refresh. [1]
March 15, 2018	Original review publication	This is the actual review date. [1]
October 6, 2018	Metashape 1.5.0 pre-release thread	The PhotoScan-to-Metashape rename came later. [9]
December 31, 2018	Metashape 1.5.0 build 7492	The rename was finalized after the March 2018 review frame. [9]
March 9, 2021	Epic acquisition announcement	Ownership changed years after the original review. [14]
June 17, 2025	RealityCapture rebrand to RealityScan	Modern naming is different from the 2018 name. [15]
July 28, 2026	Page update date	The page can look current even though the review evidence is old. [1]
August 3, 2026	Legacy licensing server cutoff date	Relevant only for current access notes, not for interpreting the 2018 test. [16]

Because the displayed models were June 2017 outputs and were not reprocessed for the later refresh, the screenshots, meshes, and impressions on the page should be treated as historical reviewer observations rather than a fresh 2026 rerun. The same caution applies to performance claims: the review does not provide a controlled metrology benchmark or a universal accuracy figure. [1]

Timeline and naming context

The naming issue matters because later labels are often projected backward into earlier software history. In March 2018, the historically correct comparison label was Agisoft PhotoScan, not Metashape. Agisoft’s own pre-release thread for Metashape 1.5.0 starts on October 6, 2018, and notes the final 1.5.0 build 7492 release on December 31, 2018, so the rename belongs to late 2018 rather than the March 2018 review frame. [9]

The broader chronology is straightforward: the review page belongs to March 15, 2018, the illustrated results belong to June 2017, Epic announced the Capturing Reality acquisition on March 9, 2021, and the product formerly known as RealityCapture was rebranded as RealityScan on June 17, 2025. If current licensing context is mentioned, the vendor notice sets August 3, 2026, as the legacy server cutoff date, which is already in the past as of September 2, 2026. [1] [14] [15] [16]

How RealityCapture turns photos into a 3D model

Photogrammetry is the practice of deriving reliable information about physical objects and environments from photographs, so the workflow here is image-based reconstruction rather than active range sensing. In this article, “scan to mesh” is only search-language shorthand. The process being discussed is photo-to-mesh reconstruction from overlapping images. [7]

Term	Meaning in this article
image alignment	Matching common features across photos to estimate camera poses; the aligned set may split into one or more components. [3]
sparse point cloud	The initial feature-based 3D cloud produced by alignment, together with camera positions. [8]
dense point cloud	A denser reconstruction stage used explicitly in PhotoScan 1.4 before mesh generation. [8]
mesh	The polygon surface reconstructed from the aligned imagery and point-cloud data. [8]
UV mapping	The unwrap stage that lays out 2D texture coordinates on the mesh before texturing. [6]
texture	The color layer projected onto the existing geometry; it changes appearance, not the underlying shape. [6]
control points	User-placed image correspondences used to tie images together or help correct pose estimation. [17]
GCPs	Ground control points with known real-world coordinates used to scale or georeference the model. [17]
check points	Ground test points kept out of optimization and used to report deviations during validation. [17]

In practical terms, the RealityCapture photogrammetry workflow starts with importing overlapping photos, then aligning them by matching shared features so camera poses can be estimated and grouped into components. After alignment, you check whether the cameras make sense, whether the scene stayed in one coherent component, and whether the sparse point cloud reflects the object that was actually photographed. Current RealityScan help is useful here as concept support: it says low overlap below 20% is risky, best quality is associated with neighboring-image overlap above 60%, and alignment precision improves when full image resolution is retained. [3]

The reconstruction stage turns the aligned photo set into fuller geometry. Current help explains the speed-detail tradeoff numerically: image downscale 1 means full resolution, while downscale 2 means each image side is halved and only 25% of the original pixel count is used. The High Detail meshing mode uses full resolution and therefore does not expose that downscale option. Historically, PhotoScan 1.4 followed the same broad stage logic: load photos, align them, generate a dense point cloud, build a mesh, texture it, and export the result, with alignment explicitly producing camera positions and a sparse point cloud. [4] [8]

After geometry comes appearance. UV unwrap and texturing help the model look complete, but they do not improve the mesh by themselves. Current RealityScan help lists export examples including OBJ, PLY, XYZ, ABC, GLB, STL, 3MF, USD, USDZ, PTX, LAS, FBX, DXF, and DAE, while the texturing help exposes resolutions from 512 to 16384 and describes a default gutter of 2 pixels. Those are useful concept and search-intent references, but they should not be treated as proof of geometric quality or automatically backdated into the March 2018 interface. [5] [6]

RealityCapture review 2018: photogrammetry workflow and results technical illustration 1 — Technical illustration for ## How RealityCapture turns photos into a 3D model.

What “accuracy” means here — and what it does not mean

RealityCapture image alignment accuracy is not a single number. A better way to think about validation is as a ladder of evidence. At the lowest level are internal diagnostics: whether alignment forms stable components, whether camera poses look plausible, and what reprojection error does in pixel space. Current help recommends a maximum feature reprojection error of 3 px and says greater overlap makes camera-pose estimation easier, but that is still internal image-space evidence. Next comes constrained scaling or georeferencing: control points, GCPs, and scale bars can anchor the project to real dimensions or coordinates. Stronger again is independent validation through check points or ground test points, because they are not used to force the solution. Strongest is external geometry comparison against reference data, where benchmark methods use ground truth and distance-threshold metrics instead of screenshots alone. Tanks and Temples, for example, defines precision and recall over a distance threshold (d) and combines them as an F-score. [3] [12] [17]

Metric	What it tells you	What it does not tell you
accuracy	Agreement with a known reference. [12]	A universal software-wide guarantee. [1]
precision	Tightness or repeatability at a stated threshold. [12]	Real-world correctness by itself. [12]
resolution	How much visual or sampled detail is retained. [4] [6]	Certified dimensional accuracy. [1]
reprojection error	Image-space fit between observed image points and projected model points, in pixels. [3]	Direct millimeter accuracy. [3]
control points	User constraints that help connect images or influence pose estimation. [17]	Independent validation by themselves. [17]
GCPs	Measured reference points with known coordinates. [17]	Automatic correctness without checking. [17]
check points	Held-out measured references for error reporting. [17]	Fitting constraints used to force the model. [17]

The 2018 review itself cannot support a universal accuracy claim because it does not publish a controlled metrology benchmark. Later papers are useful only as conditional evidence. One industrial comparison tested 6 software solutions across 4 industrial datasets captured with 2 DSLR cameras and evaluated 12 calibrated scale bars; in that specific study, RealityCapture recorded a 0.3 px reprojection error, while PhotoScan and RealityCapture both reached 0.1 mm RMSE on the scale-bar distances. A separate cultural-heritage study used 169 images and reported an average total processing time of 133 minutes for Metashape, while RealityCapture was the fastest in that test summary. Those numbers do not settle the software “overall”; they show that credible comparisons depend on dataset, capture setup, settings, hardware, and the chosen validation method. [1] [10] [11]

RealityCapture review 2018: photogrammetry workflow and results technical illustration 2 — Technical illustration for ## What “accuracy” means here — and what it does not mean.

RealityCapture vs PhotoScan in the 2018 frame

The historically correct comparison for March 2018 is RealityCapture versus Agisoft PhotoScan. Metashape belongs only in a rename note, because Agisoft’s own record places the Metashape 1.5.0 pre-release on October 6, 2018, and the final build 7492 on December 31, 2018. [8] [9]

Criterion	2018 reviewer observation	Evidence basis	What not to claim
Workflow shape	Both products followed staged photogrammetry pipelines. [1] [8]	PhotoScan 1.4 documents alignment, dense cloud generation, mesh building, texturing, and export. [8]	Do not claim that one of them skipped alignment or reconstruction.
Alignment output	PhotoScan explicitly describes camera positions plus a sparse point cloud after alignment. [8]	This is a workflow-stage description, not an external accuracy certificate. [8]	Do not translate sparse-cloud output into checked real-world accuracy.
Naming context	PhotoScan is the correct March 2018 name; Metashape is later. [9]	Agisoft’s own pre-release and final-release timing establish the rename sequence. [9]	Do not call the March 2018 competitor “Metashape.”
Results boundary	The displayed models were June 2017 models and were not reprocessed for the later page refresh. [1]	The review page itself carries that caveat. [1]	Do not describe the page as a clean 2018 or 2026 retest.

The comparison is most useful when read as a photogrammetry software comparison RealityCapture case study, not as a universal verdict. The 3D Mag review is tied to one reviewer, one workstation, one set of settings, and a results section that includes June 2017 models not rerun for the later refresh. If a dataset, protocol, or controlled measurement figure is not published for a claim, the responsible wording remains: no reliable figure found. [1]

That same historical boundary should stay attached to any result summary in this section. Whatever one concludes about speed, geometry, or texture presentation from that page, the displayed models remain June 2017 outputs rather than freshly regenerated late-2018 or 2026 evidence. [1]

RealityCapture review 2018: photogrammetry workflow and results technical illustration 3 — Technical illustration for ## RealityCapture vs PhotoScan in the 2018 frame.

Where the workflow works well — and where it struggles

The workflow is usually more comfortable on subjects that offer stable, repeatable visual features across many overlapping images. The historical review leaned positively toward statues and other organic subjects, while flatter technical objects looked less convincing to the reviewer. That matches broader capture guidance from PhotoScan 1.4, which explicitly advises avoiding not-textured, shiny, highly reflective, transparent, and absolutely flat objects or scenes, and notes that geometry must be visible from at least two cameras to be reconstructable. [1] [8]

Common failure modes include these cases. [8]

Glossy or reflective surfaces that change appearance with viewpoint.
Transparent objects that do not present stable surface features.
Weakly textured or absolutely flat areas with too little unique detail.
Repetitive patterns that confuse feature matching.
Thin gaps, occlusions, and narrow structures that are hard to see cleanly from enough angles.

Any speed reading from the historical review has to stay tied to the published machine: an Alienware Aurora R5 with an i5 dual-core CPU, 32 GB RAM, and an Nvidia GTX 1070. Reconstruction downscale changes the balance between detail and speed, while texture settings change appearance options rather than geometric truth. Later comparison work is useful only as a reminder that results remain dataset-specific: in the industrial study, for example, RealityCapture’s 0.3 px reprojection error and 0.1 mm scale-bar RMSE with PhotoScan were conditional outcomes from that exact setup, not a blanket guarantee. [1] [4] [6] [10]

Exports, textures, and what they do not prove

Export is a packaging step, not a validation step. Geometry deliverables such as point clouds and meshes describe structure, while textures and UV layouts describe surface appearance. A model can export cleanly and still have weak geometry, alignment problems, or surface noise; the reverse can also happen, where solid geometry is paired with mediocre texturing. [5] [6]

Current RealityScan help lists a broad set of output formats, including OBJ and FBX, but that modern export list should be used only for present-day documentation context. It should not be backdated wholesale into March 2018 behavior unless the older build is independently documented. The safe general point is that export support does not prove reconstruction quality. [1] [5]

What changed after 2018

Two later changes matter for context. Epic Games announced on March 9, 2021, that Capturing Reality had joined Epic Games, and on June 17, 2025, the desktop product formerly known as RealityCapture was rebranded as RealityScan. Those events change present-day ownership and naming, but they do not change what the 2018 review actually tested. [14] [15]

If current access needs a date note, the vendor notice says the legacy RealityCapture licensing server closed on August 3, 2026, and that users would lose access at that point unless they updated or converted affected projects. As of September 2, 2026, that cutoff date is already in the past. [16]

Practical reading guide for this archival review

Use the page as a historical workflow reference, not as a current ranking page. A short checklist helps keep the interpretation honest. [1]

Check the dates first: the review was published on March 15, 2018, the page later showed an update date of July 28, 2026, and the displayed result models were June 2017 outputs that were not reprocessed for the later refresh. [1]
Check the machine and settings before inferring speed: the published hardware was an Alienware Aurora R5 with an i5 dual-core CPU, 32 GB RAM, and an Nvidia GTX 1070. [1]
Check the validation method: control points, GCPs, check points, scale bars, and reference-cloud comparisons answer different questions. [12] [17]
Check the product names: March 2018 means PhotoScan in the competitor frame, not Metashape. [9]
Check whether a claim comes from the historical review or from current RealityScan documentation, because present-day help pages are useful for explaining concepts but are not direct evidence of the 2018 interface or feature set. [3] [4] [5] [6]

Bottom line — what a RealityCapture review 2018 still tells readers

A RealityCapture review 2018 page is still useful if the goal is to understand one historical workflow, one reviewer’s observations, and the kinds of outputs that impressed or disappointed that reviewer at the time. [1]

Its limits matter just as much as its value. The review was published on March 15, 2018, the displayed models were June 2017 outputs that were not reprocessed for the later refresh, and the page does not support a universal accuracy figure. Modern naming belongs in a separate bucket, because the product was rebranded as RealityScan in 2025. The article remains worth reading as archive and context, but not as a present-day ranking, pricing guide, licensing guide, or certified-accuracy source. [1] [15]

FAQ

What is the RealityCapture photogrammetry workflow?

At a practical level, it is the sequence from overlapping photos to a model: import images, align them by matching shared features, inspect components and camera poses, generate geometry, unwrap the mesh, texture it, clean it up, and export it. Current help is useful for explaining the concepts behind those stages, while PhotoScan 1.4 helps anchor the historically similar 2018 competitor workflow. The important boundary is that this article is about photo-based reconstruction, not active scanning. [3] [5] [7] [8]

How accurate is RealityCapture image alignment?

There is no single universal number. The strongest answer is layered: internal alignment diagnostics such as reprojection error are measured in pixels, not millimeters, and current help recommends a maximum feature reprojection error of 3 px. Real-world accuracy depends on additional evidence such as GCPs, scale bars, held-out check points, or reference-geometry comparisons. The 2018 review itself does not publish a controlled metrology benchmark, so it cannot justify a universal claim. [1] [3] [10] [17]

How does RealityCapture scan to mesh work?

In this article, “scan to mesh” means photo-to-mesh reconstruction. The software aligns photos, estimates camera poses, reconstructs geometry from the aligned imagery, and then builds a mesh that can be textured and exported. That is different from saying the geometry came from an active scanner. Mixed-input or laser-scan workflows are a separate topic from the historical review discussed here. [4] [5] [7]

Is RealityCapture better than other photogrammetry software in 2018?

Not in any universal sense that this review can prove. The historically correct comparison frame is RealityCapture versus PhotoScan, not Metashape, because the Metashape rename appears only later in 2018. PhotoScan 1.4 clearly used a staged workflow of alignment, dense cloud generation, mesh building, texturing, and export, so the meaningful comparison is about how two workflows behaved on a given test rather than about brand mythology. The 3D Mag page is informative, but it is not a controlled leaderboard. [1] [8] [9]

What is the difference between control points, GCPs, and check points?

Control points are manually identified image correspondences used to help tie images together or improve pose estimation. Ground control points extend that idea by adding known real-world coordinates, which lets the model be scaled or georeferenced. Check points, called ground test points in current RealityScan help, are measured points reserved for reporting deviations rather than optimizing the solution. That distinction is why a project can look internally tidy but still fail external validation if the held-out points disagree. [12] [17]

Can RealityCapture export OBJ or FBX files?

Yes, in current documentation context. Current RealityScan help lists OBJ and FBX among the available export targets, along with several other geometry and point-cloud formats. That is useful for modern search intent, but it does not automatically tell you what every 2018 build supported in exactly the same way. More importantly, successful export only tells you the data can be written out; it does not certify that the mesh is clean or metrically trustworthy. [5]

Is this RealityCapture review 2018 still useful in 2026?

Yes, if you want historical workflow context and reviewer-observed results. No, if you want a current buying guide, present-day licensing advice, or a certified accuracy benchmark. The reason is simple: the page combines a March 15, 2018 publication, a July 28, 2026 page update, and displayed June 2017 result models that were not reprocessed for the later refresh. Modern naming adds another layer, because RealityCapture was rebranded as RealityScan in 2025. [1] [15]