Last week, I was in the Netherlands to lead two workshops on the topic of Gaussian splatting and also to introduce 3D scanning. After the workshops, we consciously took an extra day to scan in Rotterdam itself. What emerged from this shows very clearly how AI and Gaussian Splatting to supplement today
For the 3D scanning in Rotterdam, we have with the PortalCam from XGRIDS The PortalCam is a LiDAR-powered camera specifically designed for rapid volumetric capture of real-world environments.
Precise classification is important here:
The point cloud generated by PortalCam is not intended as a high-precision end product. Its primary purpose is spatial orientation and alignment of the captured image data.
The LiDAR data help to:
To reconstruct camera positions in space stably
Aligning image sequences geometrically correctly
Minimize drift in longer scans
to create a consistent basis for later reconstruction
The actual visual quality of later Gaussian splats primarily arises from the image information – not from the geometric precision of the LiDAR points themselves.
With a price of around 5,000 Euros, the PortalCam is currently remarkably positioned within its category. It combines real-time visualization, integrated tracking, and workflow speed suitable for media productions in a mobile device.
We mounted the PortalCam on a long extension pole. This allowed us to scan significantly higher than would be possible from normal eye level.
This has decisive advantages:
Reduction of occlusions by pedestrians or vehicles
better capture of upper facade areas
higher probability of capturing roof edges and upper building structures
More consistent results when viewed later from elevated virtual perspectives
From the ground, it is very difficult to accurately capture roof surfaces or upper facades because the scan angle is too flat. With an elevated camera position, objects can easily be viewed from above, thereby reducing geometric gaps.
This is particularly relevant for urban scenes. When you view the finished Gaussian Splat virtually from a height above human eye level, inadequately captured areas immediately become apparent in the form of holes or blurry-looking surfaces.
The entire scan of the street block took about 30 minutes.
I don't like working alone on large urban scans. Johannes Müller assisted me in Rotterdam. While operating the PortalCam, XGRIDS' LiDAR scanner on the extension pole, you don't have a free hand to keep an eye on the XGRIDS app, LCC-Go, on your smartphone. This app shows in real-time which areas of the surroundings have already been captured by the scanner.
The app visualizes the resulting point cloud in real time. This allows you to see:
which areas have already been covered
where data is missing
whether there are potential gaps in the scan
While walking, one focuses on movement, stability, and secure camera control. This can lead to areas being unintentionally missed. These would later result in unclean or missing geometry fragments in the Gaussian Splat.
Leading the PortalCam on a long extension pole is also physically demanding. Due to the weight and value of the device, it should be held steady with both hands. A fall would not only be technically problematic but also economically significant.
The scan itself is only the first step in the 3D scanning workflow. The crucial part is processing the raw data into a photorealistic Gaussian Splat.
Upon completion of the scan, the recorded data will be LCC Studio from XGRIDS imported. The LCC Studio is specifically designed to efficiently process image and LiDAR data generated by the PortalCam.
The workflow is remarkably straightforward:
Import of Raw Data
automatic alignment of camera positions
Optimizing Image Relationships
Export as Gaussian Splat dataset
In practice, this is almost drag-and-drop. After import, the system largely automates the reconstruction. Within a short time, a photorealistic Gaussian Splat is created, which can be viewed directly in the viewer.
The technical classification is important here:
The LiDAR data primarily supports the spatial alignment of the image data. The visual quality of the finished Gaussian splats is based on the image information. LCC Studio combines both information sources into a stable volumetric model.
The dataset generated in Rotterdam was exceptionally large, 10 gigabytes. Urban environments with a lot of geometry, facade structures, and street space generate correspondingly extensive amounts of data. The amount of data increases in relation to how long you scan.
For this reason, I have decided to perform the calculation using the LCC Cloud to have performed.
Rendering on a local system would have likely taken over 30 hours – possibly even longer depending on the GPU setup. The computation could be performed much more efficiently in the cloud, and I didn't have to endure the fan noise of my graphics card for hours on end.
Cloud processing is therefore a sensible part of the workflow, especially for large-scale 3D scanning. It reduces:
local hardware load
Render times
Default risks
and enables parallel work on other projects
The result is a high-resolution Gaussian Splat that can be viewed and further processed in real-time.
A central theme in 3D scanning of urban environments is dealing with moving objects. Cities are not static spaces. Vehicles, pedestrians, and bicycles constantly change the image information during the scan.
During reconstruction, XGRIDS's software detects moving elements and attempts to remove them from the final Gaussian Splat.
In practice, this works reliably for pedestrians. People moving through the scene are correctly identified and generally do not leave significant artifacts in the model.
It becomes more complex with objects that only move intermittently. A typical example is cars at a traffic light.
If a vehicle is stationary for part of the scan, it can be interpreted as part of the static environment. If it then continues to move, so-called Floater – semi-transparent, cloud-like fragments in space.
Floaters are caused by conflicting image information within the reconstruction. Parts of an object are captured in different positions without the algorithm being able to clearly assign them to a fixed structure.
The result is diffuse, ghostly geometric fragments that appear to stand freely in space.
For cleanup, I have 3DGS tools like SuperSplat and PostShot used.
The affected areas must be manually selected and removed. This can take time. A reliable automatic floater detection does not currently exist. For large-scale urban datasets, this manual correction remains part of the regular Gaussian splatting workflow.
After cleaning, the Gaussian Splat files:
largely free of floaters (I could have spent hours on it.)
geometrically consistent
The scene could be viewed and navigated easily from different perspectives.
Nevertheless, a fundamental problem became apparent: the urban space appeared unrealistic without movement.
Especially in urban scenes, the algorithmic removal of moving objects inevitably leads to an artificially empty environment. The model correctly depicts the built structure, but not the actual use of the space.
This is where the classic 3D scanning and Gaussian Splatting process ends.
What follows is a separate design step.
The addition of people or movement is not an integral part of the Gaussian splatting method itself. It's done subsequently by processing rendered imagery of the 3D model.
Gaussian Splatting creates a volumetric 3D model.
The AI does not work directly on the point cloud or splat data, but on the video generated from it.
In the XGRIDS Viewer, a neutral avatar can optionally be displayed. By default, it is a black and white robot with branding.
For a quick test, I implemented the following workflow:
Navigating the finished Gaussian Splat
Recording a camera movement via screen capture
Importing videos into Click
Replacement of the robot with a realistically dressed person
Add additional passers-by via prompt
The 3D data remain unchanged. The AI exclusively modifies the rendered video material.
The result is a scene with visible human activity, while the structural building is entirely from the 3D scan.
Are you interested in developing a virtual reality or 360° application? You may still have questions about budget and implementation. Feel free to contact me.
I am looking forward to you
Clarence Dadson CEO Design4real
