Fifteen years into the UAV software revolution, Pix4D’s CTO sees Gaussian splatting, AI-assisted editing, and the smartphone-as-surveying-instrument converging into something bigger than any single tool.
When Pix4D was founded in 2011 as a spin-off from EPFL’s Computer Vision Lab in Lausanne, the pitch was essentially this: drones are coming, and someone has to turn the images they collect into something surveyors can use. The company grew with the UAV industry through a decade of expanding hardware, tightening regulation, and widening applications — construction, agriculture, utilities, civil engineering — until the original photogrammetry core was supporting use cases its founders likely never anticipated.
At Geo Week 2026 in Denver, Pierangelo Rothenbühler, CTO at Pix4D, sat down with xyHt to talk through where the company stands now — and what it’s building toward. The conversation ranged from Gaussian splatting’s integration into production workflows to Meta’s Segment Anything Model showing up in desktop software, to the increasingly serious proposition that a smartphone with RTK is a legitimate field instrument.
A Software Company in Three Layers
Rothenbühler is careful about how he describes the stack. Pix4D is a software company — full stop — but software here means something layered. The first layer is sensor compatibility: cameras on drones, LiDAR sensors, thermal imagers, phones. The second is reconstruction: turning that sensor data into georeferenced, accurate, reproducible representations of reality. Point clouds, meshes, orthomosaics — the outputs that have always been the company’s calling card, with precision as the non-negotiable constraint. The third layer is where the enterprise play lives: extracting insights, comparing conditions over time, running QA/QC against CAD designs, managing access and permissions across teams of hundreds.
“We’ve bridged the gap between the surveyor’s requirement for absolute accuracy and the enterprise’s demand for speed, scale, and ease of use,” Rothenbühler said, describing how large clients now use the platform to monitor multiple sites simultaneously, track construction progress against design intent, and manage who sees what within a unified cloud environment.
The word he responded to most warmly, when pressed, was “productivity platform.” Not because Pix4D is abandoning its surveying roots — Rothenbühler was emphatic that hardcore surveyors remain a core constituency — but because the cloud layer has made the underlying geospatial data accessible to a wider set of stakeholders within an organization. A project manager who doesn’t know what a point cloud is can still pull meaningful information from a PIX4Dcloud dashboard. That accessibility, without sacrificing the precision that makes the data trustworthy, is what the enterprise pitch is built on.
The utility sector has emerged as a particular growth area. Underground infrastructure — pipes, cables, conduit — needs to be mapped, recorded, and maintained in databases of record. The combination of drone photogrammetry for corridor and surface capture, smartphone scanning for trench-level documentation, and RTK integration for absolute accuracy gives utility operators a workflow that didn’t exist at this price point or field simplicity five years ago.
Gaussian Splatting: Accuracy First, Visual Fidelity Second
Pix4D has been among the more technically aggressive of the established photogrammetry vendors in its approach to Gaussian splatting — not as a novelty, but as a pipeline component. Rothenbühler described the company’s implementation philosophy in terms that distinguish it from how most of the broader visualization world has engaged with the technology.
The default Gaussian splatting implementation most practitioners encounter optimizes for visual fidelity — the photorealism that makes the technique arresting. Pix4D’s approach reverses the priority order. Geometric fidelity comes first: the Gaussian splatting model is anchored to the same photogrammetric reconstruction that generates the point cloud and mesh, meaning the splat is georeferenced in a projected coordinate reference system and measurable against real-world coordinates. Visual fidelity is layered on top of that foundation, not substituted for it.
The practical consequence is significant. Rather than offering Gaussian splatting as an alternative output alongside point clouds and meshes, Pix4D runs it as part of the reconstruction pipeline itself. The splats inform the point cloud generation — resulting in denser, cleaner point clouds, particularly in areas that have historically challenged photogrammetry: thin structures, cables, scaffolding, scaffolding pipes, glass surfaces. Construction teams relying on point clouds to answer questions like “is the wall straight” or “does this match the design model” get better underlying data, not just a more appealing visualization.
Pix4D introduced Gaussian splatting for terrestrial capture via PIX4Dcatch in 2025, presented production-ready results at Geo Week 2025, and extended the capability to drone datasets in PIX4Dcloud at Intergeo 2025. By Geo Week 2026, the company had run a dedicated session on its approach — the one Rothenbühler referenced — walking through what the pipeline integration means in practice. The session reflected a broader conference moment: Gaussian splatting, once a research preview and then a novelty feature, was being evaluated at Geo Week this year as a production tool by practitioners with real deliverables on the line.
Segment Anything: AI as a Workflow Accelerator
The Segment Anything Model (SAM), released by Meta as a foundational vision model in 2023, segments objects in images based on a point or region prompt — essentially, you click on something and it finds the best boundary. Pix4D was among the first to integrate SAM into a production geospatial workflow, embedding it directly into the desktop software.
The use cases Rothenbühler described are concrete and unglamorous in the best possible way. In orthomosaic editing, cars parked on a survey site during capture create artifacts that need to be removed before delivering the dataset. Traditionally, that’s manual work — drawing selection boundaries around each vehicle individually. With SAM integration, a single click selects the car, and the software extracts it. Applied across a large orthomosaic with dozens of vehicles, the time savings are substantial.
Point cloud classification is the other major application. Assigning semantic labels to point cloud data — distinguishing ground from vegetation from structure — is fundamental to many downstream workflows and has historically required significant manual review. SAM-assisted classification lets a user click on a region or object and have the selection propagate automatically, reducing the manual boundary-drawing that makes classification tedious at scale.
Neither application replaces the professional judgment that makes the classification meaningful or the edit accurate. What SAM does is compress the time between decision and execution, which is precisely where the productivity gains Rothenbühler describes throughout the platform conversation become tangible.
“We were doing AI even before it was a hype,” he noted — a point worth taking at face value given that photogrammetry has always involved machine learning at its core, long before the current wave of foundation models. The SAM integration is the company extending that lineage into the new toolset rather than retrofitting AI onto an existing product.
PIX4Dcatch: The Phone as Field Instrument
Perhaps the most commercially interesting development Rothenbühler raised is PIX4Dcatch, the mobile platform that the company describes as turning a smartphone into a surveying tool.
The underlying proposition is straightforward: iPhone Pro and iPad Pro devices now ship with LiDAR sensors capable of supporting close-range capture. Pix4D’s software pipeline takes video or image sequences from those devices, processes them through the same reconstruction engine that handles drone data, and delivers georeferenced outputs through PIX4Dcloud. For applications like trench documentation — capturing the as-laid position of a utility installation before backfill — the phone-based workflow is often the most practical option available. Drones can’t get into a trench. Terrestrial scanners are overkill for a linear job that needs to move fast. A phone that a field technician already carries is the right tool.
The RTK integration closes the accuracy gap that historically made mobile capture unsuitable for survey-grade work. Pix4D has formalized partnerships with Trimble, Topcon, Leica, Bad Elf, and Emlid to combine their RTK solutions with PIX4Dcatch, adding absolute accuracy in a projected coordinate system to the phone’s relative reconstruction. The result is a workflow that can produce centimeter-level positional accuracy from a device that weighs less than a pound — a combination that would have been implausible at production scale even three years ago.
The enterprise angle is that this scales. A single licensed platform can deploy PIX4Dcatch to field crews across an organization, with all captures feeding into the same PIX4Dcloud environment where drone data, desktop projects, and other data sources converge. The trench scan a technician captures in the field on Tuesday is in the same database as the site orthomosaic from the drone flight on Monday. That kind of integration, at a price point and operational simplicity accessible to utility operators and construction contractors rather than just specialist survey firms, is where Rothenbühler sees the near-term growth.
The Platform Logic
What ties these threads together — Gaussian splatting as pipeline infrastructure, AI-assisted editing, RTK-equipped smartphones — is a consistent underlying argument: the value of geospatial data scales with how many people in an organization can act on it. Pix4D’s 15-year core competency in accuracy and reliability is the foundation that makes the wider platform trustworthy. The enterprise cloud layer, the mobile platform, the AI integrations are all in service of the same objective: getting that accurate, reliable data in front of the project managers, asset managers, and field supervisors who need it, without requiring each of them to become photogrammetrists.
That’s the productivity platform frame — and it’s a description Rothenbühler, who came up through Pix4D as a technical support engineer before moving into business development, product management, and eventually the CTO role, has clearly thought carefully about. The company that grew with the drone industry is now making the case that geospatial intelligence is broader than any single capture method.