In recent years, photogrammetry has evolved significantly, giving rise to new ways of creating digital replicas through images. Depending on the objective pursued when generating these 3D models, different techniques can be applied.
The generation of digital replicas of objects, environments, buildings, industrial plants, etc., is extremely useful for study and optimisation, planning interventions and improvements, inventory management, efficiency calculations, and many other applications.
In some cases, only a visual inspection is required; in others, it is necessary to measure distances, pipe sections, surfaces, volumes, and other elements with precision. The range of needs associated with a digital replica can be as broad as the variety of clients, use cases, and application fields.
To illustrate a specific use case of digital twin creation, we can mention an example observed in one of the MOTIVATE XR pilots. The company Gorenje manufactures home appliances, specifically washing machines. The company needs to replicate its washing machines in 3D in order to program a virtual procedure that serves as a reference for maintenance and repair technicians. In this way, users can visualise different parts of the washing machine in 3D, see in mixed reality where each component should be placed, or follow step-by-step instructions through XR/AR glasses.
In this case, precise measurement may not be a necessary feature; instead, visual quality becomes more important.
Why use images instead of laser scanners or structured light sensors? The goal is to achieve a method of replicating reality that is affordable and accessible. 3D scanning systems based on laser technology or structured light sensors have a higher cost compared to simple video cameras integrated into smartphones, camcorders, drones, or XR/VR headsets. Image-based capture therefore becomes a more democratic and straightforward way of scanning, accessible not only to professionals but especially to non-professionals.
How to Create Digital Models from Images and Videos
Photogrammetry
Photogrammetry is the technique responsible for generating 3D models from images. It was developed in the 19th century, only a few years after the official presentation of the daguerreotype; the French engineer Aimé Laussedat was the first to use photographs for topographic surveying.
Photogrammetry consists of using multiple images (at least two) to calculate points in space. This mathematical operation is based on the fact that the camera projection centers, the measured ground point, and the projection of that point onto the images (free of distortion) lie on the same plane. This leads to the coplanarity condition, from which the three-dimensional coordinates of the point in space are computed.
This simplified explanation can be complemented by modern techniques that enable large-scale 3D reconstructions using depth maps and triangulation, based on the orientation of multiple images, calibrated cameras, and related processes.
Traditionally, photogrammetry was performed by capturing images one by one. However, several years ago it was observed that video recording could also provide a valid set of images to perform photogrammetry in a more automated manner.
Fig 1. Epipolar Geometry.
Two cameras take a picture of the same scene from different points of view.
Videogrammetry
This led to the emergence of videogrammetry, where the user no longer captures images individually but records a video from which an algorithm extracts the necessary frames.
At 2Freedom Imaging, we have developed a videogrammetric procedure that we have optimised over recent years, from intelligent frame selection to full 3D reconstruction from sequential captures. In this way, it is possible to generate accurate 3D models in a versatile manner using video.
The main limitation of photogrammetry (and consequently videogrammetry) lies in the search for homologous points between images with overlapping areas. In some cases, this operation is not straightforward, especially when dealing with homogeneous or reflective surfaces. In such situations, the algorithm may fail to identify common points, resulting in holes or deformations in the model.
Gaussian Splatting
The Gaussian Splatting technique requires previously oriented images, as it is a method for 3D reconstruction and representation rather than for image orientation. Therefore, this technology is linked to classical photogrammetric procedures during the initial image orientation phase.
Gaussian Splatting is a modern 3D representation and rendering technique that uses thousands or millions of 3D Gaussians (mathematical functions with an ellipsoidal shape) to reconstruct and visualise scenes with high realism. Popularised in 2023, this technology represents spatial information through semi-transparent 3D ellipsoids that store position, shape and orientation (covariance), color, and opacity.
This representation must be visualized using specific renderers, typically volumetric or GPU-based, in order to achieve a very high level of visual realism.
Gaussian Splatting viewers do not display points or meshes, but rather millions of dynamically projected ellipsoids. The user perceives a continuous scene, free from the typical noise of point clouds, with smooth transitions between different viewpoints.
Gaussian Splatting
3D Model – Lavadora test – SuperSplat
Comparison
We are not dealing with equivalent techniques. Videogrammetry and Gaussian Splatting share a common initial phase: the relative orientation of images. However, the method of 3D reconstruction and representation is entirely different.
Gaussian Splatting is clearly focused on visualisation. In fact, when analysing the exported points, one observes a high level of noise, with cloud-like structures that do not precisely define the real geometry of the object. This occurs because Gaussians are not designed to accurately define geometric position, but rather to optimise visual representation.
Fig. 2. Point Cloud:
Videogrammetry points (left): 340,475; Gaussian Splatting (right): 289,815 points.
Fig. 4. Generated Mesh:
In the 3D Gaussian Splatting model, the mesh generated from the point cloud appears as shown in the image on the right. Videogrammetry mesh (left).
Fig. 3. Textured Model Visualisation:
Videogrammetry on the left and Gaussian Splatting on the right.
Videogrammetry, on the other hand, offers higher geometric accuracy by technical definition (except in homogeneous or reflective surfaces). However, when deformations appear on poorly textured (very homogeneous) surfaces, they become visible in the model. In contrast, Gaussian Splatting may visually smooth or conceal such deformations, even though geometric error still exists.
Conclusion
With the emergence of Gaussian Splatting, there has been some confusion in the measurement market, and it still persists. In many cases, visual quality has been mistaken for geometric quality. The fact that Gaussian Splatting provides excellent visual results does not necessarily mean that its geometric accuracy is equivalent.
It is possible that in the future, thanks to artificial intelligence and new advances, geometric quality will improve significantly. However, at present, it does not reach the level of geometric accuracy provided by photogrammetry or videogrammetry when oriented toward measurement purposes.
There are LiDAR-based systems that incorporate Gaussian Splatting as a visual support layer, but they rely on LiDAR to provide reliable geometric information.
Videogrammetry is a natural evolution of photogrammetry and is primarily focused on geometric requirements. Nevertheless, as can be observed in the model presented in this publication, when there is even minimal texture heterogeneity, videogrammetry can also deliver high-quality visual results.
Acknowledgments
I would like to thank my friend Diego Ramírez (Chile), from GERTARQ (Getarq.com), for collaborating in the development of the Gaussian Splatting content included in this article.
Author
2Freedom Imaging
Pedro Ortiz-Coder, PhD in Geotechnologies, is co-founder, CEO and CTO of 2Freedom Imaging Software & Hardware S.L. An expert in videogrammetry, Visual SLAM and intelligent 3D digitisation, he leads the development of innovative hardware and software for surveying, construction and heritage. He combines academic and industrial experience, driving disruptive solutions in spatial data capture and analysis.
