3D Gaussian Splatting for Drone Operators, Explained

You’re standing in a construction site at 6 AM. The client is flying in at noon to see the latest progress. You have a photogrammetric orthomosaic and point cloud from yesterday’s flight — sharp, accurate, exactly 3.5 cm horizontal RMSE. But when you hand the client a textured mesh, it’s choppy. Jagged polygons. Texture popping. The client squints at their laptop and says, “Show me something I can actually walk through in 3D without it looking like a 2003 video game.”

That’s where 3D Gaussian Splatting comes in. Same flight data. Two hours later, you hand them a photorealistic, real-time interactive 3D walkthrough rendered at 60 frames per second. No polygon artifacts. No texture seams. Photorealistic appearance from every angle.

This isn’t NeRF hype. It’s not theoretical. DJI shipped native 3DGS in Terra v5.0 in July 2025. Pix4D integrated georeferenced splatting in Pix4Dcloud in October 2025. OpenSplat, created by Piero Toffanin, who created WebODM and remains its lead developer, trains splats straight from ODX projects, and WebODM 3.3.0 (September 2026) added a Splats dialog for exporting training data and viewing splats. Contractors, real estate firms, and heritage documentation projects are using this right now.

Here’s the bottom line: 3D Gaussian Splatting is not a replacement for photogrammetry. It’s an augmentation. Same images, same flight plan, different pipeline. One produces measurement deliverables. The other produces visualization deliverables. If you’re not offering both, you’re leaving money on the table.

Drone mapping workflows have evolved steadily for years. Since late 2024, published 3DGS pipelines have been tested on real project data — training times, accuracy, outputs, and what clients actually want. Here’s what the research papers get right, what they oversell, and what it means for your business.

What 3D Gaussian Splatting Actually Is

A 3D Gaussian Splat is an ellipsoidal blob — a 3D Gaussian distribution with position, orientation, opacity, and color. Train a model on overlapping drone images, and it optimizes millions of these blobs to reconstruct the scene’s appearance. That’s it. That’s the core insight.

The foundational paper — Kerbl et al. (2023, ACM SIGGRAPH, 3D Gaussian Splatting for Real-Time Radiance Field Rendering) describes the method clearly. You start with a sparse point cloud from Structure-from-Motion. You instantiate a 3D Gaussian splat at each point. Then you optimize: run forward passes where the splatting engine renders each Gaussian to the image plane, compute the difference between rendered and observed pixels, backpropagate errors through a differentiable rasterizer (not ray-marching, unlike NeRF), and update positions, rotations, colors, and opacities.

The differentiable rasterizer is the magic. Traditional rendering of Gaussians would be prohibitively slow. NeRF-style volumetric rendering — sampling a ray hundreds of times through a neural field — is even slower. The rasterization approach (direct splat-to-pixel projection) rendered at well over 100 fps in the original paper’s benchmarks. Mip-NeRF360, the leading NeRF method it was measured against, rendered the same scenes at 0.06-0.14 frames per second, versus 134-154 fps for 3DGS. That’s more than a thousandfold difference.

Colors are stored in spherical harmonic coefficients (not raw RGB), which means the model learns view-dependent appearance — a tree looks different from the side than from above because the harmonics capture directional shading. Point opacity is learned too, so the optimizer can make distant or occluded regions transparent rather than forcing them to solid color.

In the original paper, training to 30,000 iterations averaged roughly 27-42 minutes per benchmark dataset on an NVIDIA A6000 (26m54s on Tanks and Temples, 36m2s on Deep Blending, 41m33s on Mip-NeRF360); individual scenes vary around those averages. Preview quality — good enough to show a client — appears at iteration 7,000, roughly 5-7 minutes in. Larger drone datasets take longer. The speed advantage over NeRF is what makes interactive iteration feasible within a production workflow.

Key constraint: 3DGS optimizes visual fidelity, not geometry. It learns appearance — how light reflects, how colors vary with view angle, where transparent regions exist. The “3D” part comes from Gaussian positions, but those positions are optimized for rendering quality, not geometric accuracy. You’ll come back to this distinction repeatedly.

3D Gaussian Splatting vs Photogrammetry: A Practitioner Comparison

The internet conflates these two constantly. Here’s where they actually differ.

Photogrammetry — Structure-from-Motion + Multi-View Stereo — reconstructs geometry. You capture overlapping images, find matching features, triangulate 3D point positions using calibrated camera geometry, and build explicit point clouds. Every point has a provenance. Bundle adjustment refines camera and point positions to minimize reprojection error. The result is a metric point cloud you can measure, classify, and export to GIS/CAD workflows. Orthomosaics and DEMs are derived products.

3D Gaussian Splatting — Learns appearance through gradient descent. Start with SfM camera poses (from COLMAP or the drone), initialize Gaussians, optimize splat colors and positions to minimize rendering error. The result renders photorealistic views. The “geometry” is a byproduct — you can extract a point cloud, but it was optimized for visual fidelity, not measurement accuracy.

This isn’t a theoretical distinction. It determines what you can actually deliver.

Which one solves the client’s problem? Depends entirely on the deliverable.

• Lawyer needs evidence for a boundary dispute? Photogrammetry. Court-admissible coordinates.
• Real estate agent wants a Zillow virtual tour? 3DGS. Photorealistic, fast rendering, no polygon artifacts.
• City engineer modeling stormwater? Photogrammetry. DTM, point cloud, CRS, checkpoints.
• Construction PM showing subcontractors current progress? 3DGS. Walk through it on a phone without stitching artifacts.

The right tool depends on the deliverable, not the drone.

The Accuracy Reality — Where 3DGS Breaks Down

Marketing materials won’t tell you this: standard 3D Gaussian Splatting is not survey-grade.

Atik (2025, MDPI Sensors, “Comparative Assessment of Neural Radiance Fields and 3D Gaussian Splatting for Point Cloud Generation from UAV Imagery”) tested Splatfacto (Nerfstudio’s 3DGS implementation) alongside two NeRF methods on UAV data from two study areas, and compared the extracted point clouds to a photogrammetric (SfM-MVS) reference. Results:

• Average point cloud error: about 1.35 m at one site, about 0.2 m at the other. That’s roughly 7-135x larger than photogrammetry’s 1-3 cm baseline.
• PSNR: 28.93 dB (Splatfacto’s best, study area 2). This is a rendering quality metric (peak signal-to-noise ratio). Higher is better for visual appearance. It doesn’t translate to geometric accuracy.
• Training time: Splatfacto finished 50,000 iterations in 4 minutes; the NeRF methods took 15 minutes. 3DGS wins decisively on speed.

The accuracy gap is narrowing. GeoRefGS (Hou et al., 2026, MDPI Drones, “GeoRefGS: Towards Georeferenced 3D Gaussian Splatting from Unmanned Aerial Vehicle Platforms”) built georeferencing constraints directly into the 3DGS optimization. They reported:

• Mean planar error: 1.5 cm (best case; see below).
• Mean vertical error: 5.3 cm.
• Trajectory deviation: mostly under 1 cm.

That’s genuinely close to photogrammetry territory. But the results come from a simulated UAV flight in a Unity environment, not real field data, and the paper’s ablation and discussion give 5.4 cm planar for the full model, so treat ~1.5 cm as a best case. GeoRefGS is a 2026 research paper with no released code — you can’t buy a SaaS license for it. No commercial software has shipped GeoRefGS or anything equivalent (as of October 2026).

Until then: use photogrammetry for any deliverable where coordinates matter.

Real-World Accuracy Benchmarks from Peer-Reviewed Research

If GeoRefGS or equivalent becomes available, revisit this. Until then, standard 3DGS is visualization-grade, not survey-grade.

Software Landscape — Pricing & Capability

Open-Source, Free & Low-Cost

Nerfstudio (Splatfacto) — The most flexible 3DGS framework. Supports multiple architectures. Exports splat PLYs, point clouds, and meshes. No georeferencing. Needs an NVIDIA GPU with CUDA; the docs estimate about 6 GB of GPU memory for splatfacto and 12 GB for splatfacto-big. Free (Apache-2.0).

OpenSplat — Created by Piero Toffanin, who created WebODM and remains its lead developer. Reads camera poses and sparse points directly from ODX, OpenSfM, COLMAP, OpenMVG, or Nerfstudio projects — closest to a turnkey drone pipeline. Runs on NVIDIA, AMD (HIP), and Apple Metal GPUs, or on the CPU (much slower). Free and open-source (AGPL-3.0); a pre-built Windows version is sold to support the project.

Luma AI — No longer an option for splats. Luma suspended 3D capture uploads in November 2024 and its plans now cover video and image generation.

Postshot (Jawset) — Windows 10+ with an NVIDIA RTX GPU. Out of beta since v1.0 (August 2025). Free non-commercial tier (watermarked); Indie plan €17/month (€204/year) plus VAT. Creates splats from video or image sequences. No GCP/CRS support. Simplest UI.

Scaniverse (Niantic Spatial) — Free iOS/Android app (paid business tiers added in 2026). Builds splats on-device from the phone camera. Exports SPZ (open-source, about 10x smaller than PLY) plus PLY and mesh formats. Mobile-optimized, no drone integration.

SuperSplat (PlayCanvas) — Free, open-source (MIT) browser-based editor and viewer. Upload a PLY or SPLAT file, edit splat visibility in the browser, export compressed versions (PlayCanvas puts its SOG format at about 15-20x smaller than PLY). Good for post-processing and sharing.

Commercial & Semi-Commercial

DJI Terra (v5.0+) — 3DGS added in v5.0, July 2025. Native 3DGS from visible-light drone imagery; the July 2026 update (v5.3) added cluster processing for 3DGS. Requires a Flagship, Cluster or Education licence (Flagship is roughly $3,700/year in the US; the Standard licence doesn’t include it). Outputs: 3DTiles, PLY splats, GeoTIFF (appearance-based, not geometrically rigorous), plus DJI’s TerraGS format.

Pix4Dcloud — Introduced georeferenced 3DGS for drone datasets at Intergeo 2025 (October 2025). Splats are anchored to a coordinate reference system through Pix4D’s georeferenced reconstruction; Pix4D doesn’t describe GeoRefGS-style accuracy constraints. Cloud processing (no local GPU). Drone splatting is charged in processing credits based on image count and megapixels. Advantage: georeferenced splats in a shipping product. Cost: processing drone datasets into splats in PIX4Dcloud requires the Pro plan, $4,490/year. Splatting is also available in PIX4Dcatch and, since May 2026, in PIX4Dmatic with a Pro licence.

Esri ArcGIS Reality — November 2025 release. Splat creation in Drone2Map 2025.2, ArcGIS Reality for ArcGIS Pro, and ArcGIS Reality Studio 2025.2; output is a 3D tile layer. ArcGIS Pro 3.6 can display Gaussian splat layers; creating them requires an ArcGIS Reality product. Esri doesn’t publish list prices for these. Limited 3DGS documentation; marketed primarily for photogrammetry.

KIRI Engine — Splatting is a Pro feature: $17.99/month or $79.99/year. Mobile and web app with cloud processing. Creates splats from video or images. Simple workflow.

The Hardware You Need

GPU requirements are modest compared to NeRF or heavy photogrammetry processing.

Minimum: An NVIDIA GPU with CUDA for Nerfstudio. Its docs estimate about 6 GB of GPU memory for standard splatfacto and about 12 GB for splatfacto-big, so a 12 GB card such as an RTX 3060 is a sensible floor.

Recommended: A 24 GB card (RTX 3090/4090 class) for large drone scenes. For reference, the original paper’s 30K-iteration runs averaged 27-42 minutes per benchmark dataset on an A6000.

Also works: OpenSplat supports AMD GPUs via ROCm/HIP and Apple Metal on M-series Macs, and can even run on the CPU (about 100x slower). Rendering is so fast that even modest hardware achieves 60+ fps — the GPU bottleneck is training, not inference.

No enterprise hardware needed. A used RTX 4090 does the job.

The Hybrid Workflow — One Flight, Two Deliverable Tracks

Nobody is talking about this yet, but the business model is straightforward.

Shoot one flight with standard mapping parameters: 80% forward overlap, 60-70% side overlap, GSD better than 5 cm. From the same images, run two parallel pipelines:

1. Photogrammetry pipeline (Pix4D, Metashape, WebODM) — point cloud, orthomosaic, DSM, checkpoint-validated RMSE. Takes 4-12 hours depending on software and image count. Deliverable accuracy: 1-3 cm horizontal, 3-5 cm vertical.

2. 3DGS pipeline (OpenSplat, Nerfstudio, Pix4Dcloud, DJI Terra) — interactive 3D visualization, photorealistic real-time rendering. Tens of minutes on a high-end GPU: the original paper’s 30K-iteration runs averaged 27-42 minutes per benchmark dataset on an A6000, and large drone datasets take longer. Deliverable: photorealistic walkthrough.

Same images. Two products. Different revenue stream.

Price the photogrammetry at your usual rate; that’s the base. This article proposes adding 3DGS for +50% (1.5x the base rate) — faster training, lower compute cost than mesh generation, attractive to clients wanting both measurement and visualization. See the pricing section below for a sourced reference rate.

Workflow integration:

• OpenSplat + WebODM: WebODM’s ODX engine generates the camera poses and point cloud (via OpenSfM). OpenSplat reads an ODX project directly, or you can open the task’s Splats dialog (WebODM 3.3.0+) and click Download Training Data, which packages the images, camera poses, and a subsampled copy of the task’s point cloud (georeferenced_model.laz, about 125k points) in COLMAP format. Train, then use Upload New in the same dialog and View in 3D.
• Pix4Dcloud: Upload images, run photogrammetry processing, add 3DGS job from the same image set. Both pipelines execute in parallel in Pix4D’s cloud infrastructure.
• DJI Terra: Import images, choose Gaussian splatting reconstruction, run. Native support; requires a Flagship, Cluster or Education licence.
• Manual: Export camera poses from any photogrammetry software (Metashape 2.1.3+ can export cameras in COLMAP format), import into Nerfstudio, train locally.

You’re not inventing anything new here. The tools exist — you’re just sequencing them.

What 3D Gaussian Splatting Cannot Do

Here’s what 3DGS will not do, regardless of your GPU or flight data.

Survey-grade point clouds. Atik found 0.2-1.35 m average error against a photogrammetric reference. That’s roughly 7-135x photogrammetry’s 1-3 cm. You cannot use standard 3DGS splats for staking out property lines, design verification, or legal disputes. GeoRefGS narrows this gap, but it’s research software, not production.

Accurate orthomosaics with ASPRS validation. DJI Terra outputs a GeoTIFF, but it’s appearance-based — the orthomosaic is a projected rendering of the splat colors, not a geometrically rigorous orthorectification. Positional accuracy is not validated. ASPRS Positional Accuracy Standards have no methodology for neural rendering outputs. If your deliverable requires ASPRS compliance, use photogrammetry.

Bare-earth DEMs or ground classification. Splats have no concept of “ground” versus “vegetation” versus “building.” They optimize appearance globally. You cannot generate a DTM from 3DGS. You cannot classify points as ground, vegetation, structure. LiDAR + photogrammetry are your tools for this.

Vegetation penetration. Unlike LiDAR, 3DGS does not penetrate canopy. If the leaves block the camera, the algorithm sees canopy top, not ground. Use photogrammetry in open areas, LiDAR in vegetated corridors.

Legal/defensible deliverables. No court, engineer, or regulatory body has established standards for 3DGS outputs. There is no certification pathway, no error traceability, no independent validation standard. Photogrammetry has ASPRS. LiDAR has USGS guidelines. 3DGS has… research papers. Until that changes, keep 3DGS in the visualization bucket.

Real-World Use Cases Where 3DGS Wins

Some of these are shipping products; others are where the technique fits best.

Construction Progress Documentation

Gauzilla Pro builds 4D Gaussian splat time-lapses from drone, phone, or 360° video, and has added BIM import for comparing design models with site progress. Weekly drone flights produce interactive 3D models. Subcontractors view the current state in a browser without stitching artifacts or polygon rendering issues. Superintendents walk through the model on a tablet at the site.

Pricing: the 1.5x add-on proposed below, with recurring revenue from weekly flights.

Real Estate Virtual Tours

Zillow launched SkyTour in July 2025: interactive aerial tours built with Gaussian splatting from drone footage, available on eligible Showcase listings shot through its premium media package. Realtor.com announced FlyAround on October 31, 2025, but it’s a different approach: its aerial views are built from Google 3D Maps satellite data, not drone captures or splats. Gaussian splatting renders tours like SkyTour efficiently on desktop and mobile without the bandwidth overhead of video.

Pricing: interactive 3D sells as a premium tier (Zillow offers SkyTour on eligible Showcase listings shot through its premium media package), so price it above your standard listing package rather than per acre.

Heritage Documentation

The Smithsonian has presented its use of Gaussian splats for historical presentation (Khronos glTF webinar, April 2026). The benefit: capture visual fidelity without aggressive photogrammetric mesh processing that can introduce artifacts on delicate surfaces.

Peer-reviewed work (Khedr et al., npj Heritage Science, 2025) compared photogrammetry, integrated laser scanning–photogrammetry (the accuracy benchmark) and Gaussian splatting on a damaged statue of a Ptolemaic queen. The integrated method was most accurate; Gaussian splatting offered sufficient accuracy and was the best fit for rapid, resource-efficient visualization and web deployment.

Government and cultural institutions have budget for specialized documentation. 3DGS becomes a consulting service with premium pricing.

Infrastructure Inspection

Cables, scaffolding, glass facades — surfaces that break traditional multi-view stereo because they are transparent, reflective, or lack texture. 3DGS handles these implicitly. The model learns that glass is transparent and renders it correctly.

Use case: bridge inspection. Cable-stayed structure over a river. Photogrammetric mesh would show texture seams where the algorithm struggled with specular geometry. 3DGS renders the cables and suspension structure cleanly. Inspector can zoom and rotate to examine all angles without polygon artifacts. Easier to spot damage.

Construction Site Security & Progress Verification

3DGS 3D models serve as time-stamped visual proof of site conditions. Contractor submits weekly 3D model as part of progress documentation. Client can revisit the interactive model to verify milestones without site visits. Dispute resolution: “Here’s the exact state on April 15” rendered interactively.

Training 3DGS on Real Drone Data — What You Actually Need to Know

Research papers gloss over this part. Here’s what matters in practice.

Flight requirements: Same as photogrammetry — 80% forward overlap, 60-70% side overlap, sub-5cm GSD. Oblique angles matter more for 3DGS because the optimizer needs multi-angle coverage to learn appearance. Vertical surfaces especially benefit from side-facing imagery.

Image quality: Standard drone JPEG is fine. No need for RAW. No need for lossless compression. The optimizer converges on appearance, not on preserving image detail.

Camera poses: Critical input. 3DGS needs accurate positions and orientations from SfM. COLMAP generates these from images. Or import extrinsics from your drone’s flight log. Or export poses from photogrammetry software (COLMAP format from Metashape, or WebODM’s training-data export).

GCPs are optional for standard 3DGS. They improve absolute position accuracy for GIS alignment but don’t change rendering quality. Use GCPs if the splat cloud needs to match existing GIS datasets or if you’re running the hybrid workflow.

Software workflow (OpenSplat + WebODM example):

1. Fly grid pattern plus obliques, 80% overlap, 200-500 images
2. Upload to WebODM and process the task (ODX runs SfM with OpenSfM)
3. Click the task’s Splats button, then Download Training Data in the dialog (a .zip with the images, camera poses, and a subsampled ~125k-point copy of the task’s point cloud, in COLMAP format)
4. Extract the .zip and run OpenSplat on the folder from the command line, e.g. opensplat /path/to/training-data -n 30000
5. Training runs on your GPU. For a quick check, run a short pass first (fewer iterations) before committing to the full run.
6. OpenSplat writes splat.ply
7. Back in the task’s Splats dialog, click Upload New to add splat.ply, then View in 3D, or open the PLY in SuperSplat. Done.

Nerfstudio workflow (local GPU required):

1. Images on disk
2. ns-process-data images --data /path/to/images --output-dir /path/to/processed
3. ns-train splatfacto --data /path/to/processed
4. Real-time rendering viewer opens in browser during training
5. Export the splat PLY: ns-export gaussian-splat --load-config /path/to/config.yml --output-dir /path/to/exports

Neither workflow is complicated. Five to ten steps, start to finish.

Export Formats & Interoperability

Know these formats so you can plan your downstream workflow.

PLY (Polygon File Format) — Universal, uncompressed. Stores each Gaussian’s position, scale, rotation, opacity, and spherical-harmonic color (62 values per splat in the standard layout). File size: 500 MB–3.7 GB (5-15 million splats). Every tool exports it. Use as archival format — readable decades from now.

SPLAT (.splat) — Binary format for web viewing. About 8x smaller than PLY (32 bytes per splat versus 248). Drops the higher-order spherical harmonics, keeps basic Gaussian parameters and base color. Slightly lower visual quality, massive size reduction. Good for web delivery and client portals.

SPZ (Niantic) — Open-source (MIT) compression format. About 10x smaller than PLY while keeping spherical harmonics (quantized). Mobile-optimized. Gaining traction for mobile AR.

3DTiles — Cesium-compatible geospatial format with LOD hierarchies. DJI Terra outputs this. Web-based viewing without local downloads. Supports multiple splats at different resolutions. Standard for geospatial web.

glTF KHR_gaussian_splatting — Khronos standardization of Gaussian splatting in glTF. Release candidate February 2026, ratified September 2026. This will be the “PDF of 3D” — standardized, broadly-supported interchange format. Early adopter: Cesium (CesiumJS since v1.133, September 2025; Cesium for Unreal v2.24.0, March 2026).

Export PLY for archival. Export SPLAT for web delivery. Export glTF-KHR (when supported) for long-term interoperability.

The Service Pricing Model

As of April 2026, no US firm offers “georeferenced 3DGS deliverables” as a named service. That’s a gap worth filling.

Market pricing reference: Drone mapping prices vary widely by region, site size and accuracy tier, and operators quote both per site and per acre. For scale, one published rate card (The Florida Drone Company, March 2026) starts a standard orthomosaic at $500 per site for up to about 10 acres, quotes larger sites per acre, and prices volume calculations and other deliverables as add-ons. Your own rate card is the right base.

Proposal: Hybrid deliverable package.

• Photogrammetry package (point cloud + orthomosaic + DSM): your base rate
• Add 3DGS interactive walkthrough: +50% (1.5x the base rate)

Illustration only: on a site you would quote at $500, the bundle comes to $750. The +50% is this article’s proposal, not market data; test it against your local market.

The math: photogrammetry dominates processing time (4-12 hours). 3DGS adds tens of minutes of GPU time (27-42 minutes on average per benchmark dataset in the original paper; large drone datasets take longer). From the same images, marginal cost is minimal (GPU electricity + storage). Pricing at 1.5x base captures added value without inflating costs.

Marketing angle: “Same flight, two deliverables. Orthomosaic for your GIS. Interactive 3D for your stakeholders.”

Where This Is Going

glTF KHR_gaussian_splatting — ratified September 2026. Khronos has standardized splatting as a glTF extension. Inflection point. With a standard format, viewers and engines have one target to support.

GeoRefGS production release. A vendor (likely Pix4D, DJI, or a startup) will ship GeoRefGS or equivalent (inline georeferencing constraints). Once accuracy hits 1-2 cm horizontal, 3-5 cm vertical — competitive with photogrammetry — the use case shifts. You’d run 3DGS for both visualization and GIS-ready point clouds. Accuracy-critical work stays photogrammetry. Visualization-primary goes 3DGS. Photogrammetry becomes optional for visual-only sites.

WebODM + OpenSplat integration. This one has partly arrived. WebODM 3.3.0 (September 2026) added a Splats dialog: export training data from a processed task, train in OpenSplat or other compatible software, then upload and view the splats inside WebODM. Training still happens outside WebODM, so it’s not yet one click.

DJI Terra keeps expanding. The July 2026 update (v5.3) added cluster processing for 3DGS.

Commercial platforms convergence. Pix4D already ships splatting (PIX4Dcatch, PIX4Dcloud and, since May 2026, PIX4Dmatic with a Pro licence), DroneDeploy has Splats in beta for Aerial Pro, and Metashape hasn’t added it as of 2.3.2 (August 2026) or the 2.4.0 preview. By year-end 2026, expect 3DGS standard alongside photogrammetry — the way all now offer orthomosaics.

Splat-specific hardware optimization. Current: generic GPU rendering. Future: hardware makers optimize splat rasterization at driver level (NVIDIA OptiX, AMD RDNA3). Rendering hits 120+ fps. Mobile support improves dramatically.

FAQ

Q: Can I generate orthomosaics from 3DGS splats?

A: Only by rendering splats to orthogonal projection — visual orthomosaic, not geometrically rigorous. DJI Terra does this, but it’s appearance-based texture, not measured DTM. For engineering/legal/regulatory use, photogrammetry wins. For visual reference and client communication, 3DGS-derived ortho works.

Q: What’s the resolution of a splat cloud?

A: Measured in point density. Typical drone 3DGS produces 5-15 million splats across 50-100 acres — roughly 50K–300K per acre (depends on altitude and GSD). Rendering resolution is independent. Display all 15 million at 60 fps because the rasterizer is efficient. Photogrammetry produces 150-500 million points, but visualization requires mesh decimation and LOD.

Q: Do I need GCPs for 3DGS?

A: No. 3DGS trains without GCPs using SfM camera poses from COLMAP or flight log. GCPs improve absolute position accuracy — splat cloud registers to GIS coordinates. Without them, it’s in arbitrary local coordinates. For visualization, GCPs don’t matter. For hybrid workflow where 3DGS aligns with photogrammetry, pass the same GCPs through both.

Q: Can 3DGS handle video input?

A: Yes. Video is densely sampled imagery. Extract frames (every Nth frame), feed to 3DGS. Dense viewpoint coverage improves training — more angles constrain where Gaussians live. Especially good for interior walkthroughs (real estate, construction), where camera motion is smoother than discrete waypoints.

Q: What’s the biggest limitation of 3DGS right now?

A: Lack of native georeferencing in production tools. Standard 3DGS trains in arbitrary local coordinates. Georeferencing requires manual pipeline (export SfM poses, apply similarity transform, hope it held). GeoRefGS solves this but is research software. Once production tools ship with GeoRefGS or equivalent, it becomes viable for measurement work. Until then, keep it visualization-only.

Q: How does 3DGS perform on oblique/nadir-only imagery?

A: Nadir-only is suboptimal. No multi-angle constraints on vertical surfaces. Buildings, cliffs, facades reconstruct poorly — flattened or distorted. 3DGS needs oblique angles. Standard grid + corner patterns work well. If your client demands nadir-only, use photogrammetry.

Q: Can I use 3DGS for aerial video stabilization?

A: Not directly. 3DGS produces a static model. But you can render the camera path through the splat field to generate stabilized video — an emerging post-processing workflow, not mainstream yet, but technically feasible.

Q: How does 3DGS handle dynamic objects?

A: Poorly. 3DGS assumes static scenes. Moving objects create “ghosts” — semi-transparent artifacts where the optimizer averages across positions. DroneSplat (CVPR 2025) addresses this with dynamic filtering, but production software hasn’t integrated it yet. For now, remove moving vehicles and people from the image set or accept the artifacts.

Q: Is there an open-source georeferenced 3DGS pipeline yet?

A: Not shipping. GeoRefGS has no released code. Practical approach: photogrammetry SfM → export poses (WebODM’s training-data export or a COLMAP export) → feed to OpenSplat or Nerfstudio → splats inherit georeferencing from SfM poses (optimized with GCPs). Workaround, not native solution.

3D Gaussian Splatting is real, fast, and producing photorealistic output that clients love. It’s not replacing photogrammetry — it’s augmenting it.

Same flight, two pipelines. Photogrammetry for measurement deliverables: point clouds, orthomosaics, DEMs with CRS and accuracy statements. 3DGS for visualization: interactive 3D walkthroughs, construction progress, real estate tours.

The business case writes itself — offer both from the same imagery at a 1.5x bundle rate. Photogrammetry dominates your processing time anyway. 3DGS adds tens of minutes of GPU time and negligible marginal cost. Clients get precision measurement and photorealistic communication from one flight.

The accuracy caveat holds: standard 3DGS is not survey-grade. Atik (2025) documented 0.2-1.35 m average error against a photogrammetric reference. GeoRefGS reported 1.5 cm planar error in simulation (5.4 cm for the full model in its ablation), but that’s still research with no released code. Use photogrammetry for coordinates. Use 3DGS for visualization.

Hardware: an NVIDIA GPU with 12 GB of VRAM (RTX 3060 class) or better. Rendering is real-time on consumer GPUs. For reference, the original paper’s 30K-iteration runs averaged 27-42 minutes per benchmark dataset on an A6000; large drone datasets take longer.

Software is fragmented but workable. OpenSplat reads ODX projects and WebODM’s exported training data directly — best fit for drone workflows. Pix4Dcloud has commercial support and cloud processing. DJI Terra has it built in (Flagship, Cluster or Education licence). Nerfstudio is the flexible open-source option.

glTF KHR_gaussian_splatting was ratified in September 2026. As support spreads, splat viewing becomes as ubiquitous as PDF. Export to this format when your tools support it.

Nobody is advertising “hybrid photogrammetry + 3DGS deliverables” as a named service yet. The pipeline exists, the software works, the client value is obvious.

Go build it.

References

Research and standards cited in this article.

• Atik, M. E. (2025). Comparative assessment of neural radiance fields and 3D Gaussian splatting for point cloud generation from UAV imagery. Sensors, 25(10), 2995. doi:10.3390/s25102995
• Hou, J., et al. (2026). GeoRefGS: Towards georeferenced 3D Gaussian splatting from unmanned aerial vehicle platforms. Drones, 10(3), 195. doi:10.3390/drones10030195
• Kerbl, B., et al. (2023). 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 1–14. doi:10.1145/3592433
• Khedr, M., et al. (2025). 3D visualization of damaged statues using Gaussian splatting and web interface integration. npj Heritage Science, 13(1), 587. doi:10.1038/s40494-025-02063-5
• Tang, J., et al. (2025). DroneSplat: 3D Gaussian splatting for robust 3D reconstruction from in-the-wild drone imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 833–843). doi:10.1109/CVPR52734.2025.00086

Related Reading

• Photogrammetry & NeRF comparison
• How to create a georeferenced NeRF from drone imagery
• Drone mapping vs. land surveying: legal boundaries
• Nerfstudio documentation
• OpenSplat GitHub
• DJI Terra product page

Legal Note

3DGS outputs that contain measurable spatial data — georeferenced splats, exported point clouds, projected orthomosaics — may trigger state surveying license requirements under the same framework as photogrammetric deliverables. The 4th Circuit’s 2024 ruling in 360 Virtual Drone Services v. Ritter upheld North Carolina’s surveyor licensing law as applied to aerial maps and 3D models “containing measurable data” (the Supreme Court declined review in April 2026), and nothing in that reasoning depends on the technology used to create them. See Drone Mapping vs. Land Surveying: Where the Legal Lines Are.