How accurate are drone volumetric analysis tools?

Our analysis provides decision-grade accuracy suitable for earthwork billing and cut/fill calculations, preventing over-ordering.

Gaussian Splatting vs. Point Clouds for Construction

The S20 SLAM scanner puts a number on the screen and you trust the number. The same scan, rendered as a point cloud, looks like static to anyone who hasn't been trained to read it. Render the same data as a Gaussian splat, and the church interior shows up — walls, pews, stained glass, the whole room — and the client can walk through it on her phone.

That's the pitch. Here's the field story behind it.

What Happened at the Church

The testing location for the SHARE S20 is a church interior in the Denver metro — same spot, walked dozens of times for repeatability. On the 2026-04-09 DMA call, the S20 point cloud came up on screen "pretty much unedited." It looked good at a distance. Zoom in and the stained-glass windows resolved cleanly. The room was recognizable.

The S20 software threw a distance readout on the screen — 3.036 m. The operator hadn't yet pulled a tape measure to confirm the number. That was on the to-do list, not because the number was suspect but because every measurement a SLAM scanner prints deserves an independent check before it goes near a contract.

Then came the splat. Same scan, different representation. The Gaussian splat from the same walk lit up the room. The reaction from a peer on the call was immediate: "Freaking amazing."

The reaction from a client, the next time this comes up in a board meeting, is the same. That's the whole reason to ship the splat.

What a Splat Actually Is

A Gaussian splat is a different way to represent a 3D scene from photographs. Instead of triangulating a sparse-to-dense set of geometric points, you train a model that places thousands to millions of translucent 3D ellipsoids — "splats" — across the scene, each carrying color, opacity, and orientation. Rendered together, the splats look photorealistic. Walk through them on a phone, in a browser, without specialized software. The model knows where the ground is, where the wall is, where the camera was when each frame was taken.

The same photographs that produce a point cloud can produce a splat. The input is identical. The output looks like a video-game version of the job site.

The catch: a splat is a visualization, not a measurement tool. The geometry inside the ellipsoids is fuzzy by design — optimized to look right at a glance, not to be queried for distances. If your client wants to know the distance from a column face to a property line, pull the point cloud. If your client wants to see what the site looked like at 11 AM Tuesday, send the splat.

Pro tip of the day: A splat is a confidence layer on top of your measurement deliverable. Same data, better persuasion. The point cloud says "this is what the site is, to the centimeter." The splat says "here is what the site looks like, so you can see what the point cloud already proved." When the client walks through the splat and signs off on the budget, the splat did its job. The point cloud did the work.

Why the Client Says Yes Faster

Two things happen in a board room when a non-technical stakeholder sees a Gaussian splat for the first time.

First, they recognize the site. The owner knows her building. The owner rep has walked the loading dock. The lender has driven past the site. The splat looks like the place they've already been, on a screen they can navigate themselves.

Second, the question changes. With a point cloud, the first thirty seconds is always "what am I looking at." With a splat, the first thirty seconds is "is this right." The second question is a budget conversation. The first is a software tutorial. Splats skip the tutorial and get you to the budget.

A peer operator put it bluntly on the same DMA call: "From a customer's perspective, the splat is ten times better than a point cloud, because customers can walk through and understand it, even if they recognize it is not computer graphics."

That "ten times better" is not a technical claim. It's a perception claim. The data is the same. The reception is different. Both are real.

The Tradeoffs That Don't Make the Marketing Slide

Splats are not free. Three things the field tells you that the demos don't.

Processing time is real, and it's not uniform. On the same DMA call, the SHARE S20 splat from the church interior took 33.5 hours to process on a recent job site. Other operators report overnight runs as the norm for typical scenes. Pix4D has a beta Gaussian splat module that's faster in some cases; the open-source 3DGS pipelines (Brush, PostShot, the Inria research stack) range from minutes for a small object to days for a large commercial site. Budget GPU time the same way you budget flight time: it's a line item, not a free add-on.

Software doesn't always show you both at once. The SHARE S20's own viewer has a known limitation: turning off the point cloud layer can hide the splat as well, and some configurations won't display the splat file at all. A Super Splat workaround exists, but the point cloud won't reload into the external viewer after a switch. Plan to host the splat outside the scanner's native software — Cesium ion, Polycam, Splatware, or a self-hosted viewer. Don't promise a client a single-app experience if the splat lives in a different stack from the point cloud.

Splats lie about reflective and moving surfaces. Windows smear. Workers fragment. Cars either blur or duplicate. Point clouds have similar problems, but the failure mode is more honest: a missing point is a missing point. A smeared splat looks like a clean surface with a soft edge. Tell your client which surfaces in the splat are visual-only, and tell them in writing.

The Workflow That Actually Works

The smart play is not "splats vs. point clouds." It's "splats and point clouds, with a clear handoff rule."

The order that holds up on real projects:

Process the flight or scan in your measurement pipeline as you normally would. GCPs marked (or SLAM trajectory verified), bundle adjustment checked, independent checkpoints measuring the result. Standard rigor.
Export the point cloud for the engineering deliverable. .las or .e57, georeferenced, with the processing report. This goes to the VDC team and the surveyor.
Train a splat from the same image set or scan in a tool that fits the dataset. Pix4D's beta splat module for drone flights. Brush or PostShot for SLAM scans. The training is GPU-bound — minutes to hours for small scenes, overnight to 33+ hours for full interior walks.
Host the splat on a platform the client can open in a browser. Polycam, Splatware, Cesium ion, or a self-hosted viewer if you're already running one.
Deliver both with a one-page note that says, in writing: "The point cloud (.las) is the measurement deliverable. The splat is the visualization. They are produced from the same images and the same control. They will agree to within the visual approximation of the splat; the point cloud is authoritative for any number you put in a contract."

That last paragraph is the part that protects you. The day a client pulls a measurement off a splat and writes it into a change order, you want a written record of which deliverable is authoritative.

When to Ship the Splat — and When to Skip It

Ship the splat when:

The buyer is a non-technical stakeholder — owner, owner rep, board, attorney, lender
The deliverable supports a budget approval or a change-order review
The site is complex enough that a 2D ortho can't communicate the layout
You're delivering to more than two people and at least one of them won't open a .las viewer
The client specifically asks "can I see it" — that question is the splat's whole reason to exist

Skip the splat when:

The deliverable is internal-only and the buyer is technical (VDC, surveyor, engineer)
The site is small enough that an ortho and a 3D mesh will do the same job
The timeline doesn't allow for splat training (a few hours minimum, overnight for most real sites, 33+ hours for full interior SLAM walks)
The client's question is a measurement, not a comprehension
The scanner or processing pipeline has a known viewer issue (the S20's host-software limitation is a current example) and the splat would have to live in a separate stack from the point cloud

The Honest Tradeoff

Gaussian splatting is a visualization technology, not a measurement technology. The mistake most providers are about to make is selling splats as a replacement for point clouds. The mistake most clients are about to make is treating splats as authoritative for numbers.

The right framing: a splat is a confidence layer on top of the same data. The point cloud says "this is what the site is, to the centimeter." The splat says "here is what the site looks like, so you can see what the point cloud already proved." When the client walks through the splat and signs off on the budget, the splat did its job. The point cloud did the work.

The day your client says yes in five minutes after a thirty-second scroll through a splat is the day the splat pays for itself. It's also the day your point cloud needs to be exactly right, because the splat just invited her to ask follow-up questions that only the point cloud can answer.

Pro tip of the day: If your client opens a splat, scrolls around for thirty seconds, and then asks you for a measurement — that's not a failure of the splat. That's the splat working. It moved the conversation past "what is this" and into "what do we do." The point cloud handles the second question. Send both, and tell the client in writing which one is the source of truth.

Gaussian Splatting vs. Point Clouds: A Church, a $5,600 Scanner, and the Client Who Approved the Budget