Gaussian Splats - A Game Changer for Drone Surveys?

Drone photogrammetry has a flaw - especially in winter. Photogrammetry has become an invaluable tool for the land surveyor conducting surveys with drones. However, particularly during the winter after the leaves have fallen, trees often appear merely as a series of floating artifacts in the final mesh model. Ortho images can help establish the spread of a tree's branches, but in these situations, we've been unable to measure the tree's height or the volume of its canopy - at least without using expensive drone LiDAR systems. That was until just recently.

What is Gaussian Splatting?

Gaussian splatting is a digital imaging technique used in 3D modelling. Unlike photogrammetry, which stitches together multiple photographs to create a 3D mesh model, Gaussian splatting works by overlaying 'splats', which are individual 3D data points floating in space that are very close together. Each splat is shaped like a Gaussian curve (like a bell curve), and they blend together to appear as if they form a seamless 3D surface. This method is particularly useful in creating detailed and accurate models of objects, especially when dealing with complex surfaces or textures.

Splat! The Game is Changing

The invention of forming 3D models from #Gaussian Splats has caused something of a stir on the photogrammetry scene. It provides fast rendering and can display fine details such as those twigs and small branches that photogrammetry has always struggled with. So, how helpful can Gaussian Splatting be to the surveyor, and how can we incorporate it into our workflow?

Data Capture - Nothing New To Learn

In the example below, we have data from a small topographical survey carried out using a terrestrial laser scanner. In addition, a drone was used to capture overlapping still photographs of the site - and a fly-around video was recorded of the site and its surroundings. The scan data was processed onto GNSS control in #Trimble #Realworks, and the still photography was processed using Bentley's #ContextCapture to form a photo-realistic mesh model and a point cloud - all on the same control network.

Creating the Gaussian Splat Model and Its Point Cloud

The drone video footage was uploaded to the Poly.cam website and processed using the Gaussian Splat option. The model was formed in around half an hour and then downloaded as a .PLY point cloud file.

The point cloud is not to scale or georeferenced at this stage, so it requires some further processing. It was then imported into #CloudCompare software where it was scaled and registered to the accurate laser scan data. The Gaussian Splat data is very noisy, but it isn't too difficult to locate good points to use for the registration process. Although this was just an exercise to prove the concept (the laser scanner had already captured the tree information missed by the photogrammetry), a point cloud of the tree canopies was then created that could be imported into our #N4CE software to help draw up the topographic survey.

The Results

These images show a section taken through a large oak tree that was surveyed in three ways - laser scanning, photogrammetry and Gaussian Splatting.

• Image 1: Shows a section through the tree in the point cloud exported from the ContextCapture photogrammetry mesh model. We can see that very little of the tree has been captured.
• Image 2: Shows the point cloud from the terrestrial laser scanner (in red) overlaid onto the one from the photogrammetry model (in purple). The point cloud from the laser scanner has captured the entire tree canopy.
• Image 3: Shows the point cloud created by Polycam from the Gaussian Splats (in green). It's a sparse and noisy point cloud, but it captures the size and height of the winter tree canopy perfectly for drawing up standard topographic surveys. It seems that large amounts of noise are present where solid surfaces appear in the Splat model. As can be seen from Image 3, that noise does tend to lie beyond the surface rather than in front of it, so it isn't too difficult to work out where the true surface should be. Fine detail seems to be much cleaner. Also, in Image 3, there is a short horizontal line just to the left of the tree. That is an overhead electricity cable - again, it coincides very closely with the accurate laser scanned point cloud.

Conclusion

Although Gaussian Splatting is only in its infancy, it's clear it can be used right now to assist the surveyor using drone-based photogrammetry to improve results. The splat data obviously lacks precision still, but combined with very fast data collection via video (although still photography can just as easily be used), it can be an extremely useful tool in the surveyor's arsenal - particularly at this time of year when the trees are bare and in other situations where photogrammetry struggles to get a full result.