A gaussian splat discussion, part 1 (probably)

Updated with graphics at bottom

Apologies in advance, this is unedited stream of consciousness 🙂

OK, we have talked about 3DGS for a while now on the podcast, from news items last year to full discussions in June (I think). For the most part, I have stuck with traditional structure-from-motion/close-range photogrammetry in my day to day processing and use. I had students do collection using Scaniverse, RealityCapture (now RealityScan), or similar app to get a sense of what could be done, but I didn’t really use it.

But, things changed as processes became more streamlined and conversations continue to grow. We at VerySpatial have always leaned a little toward the geovisualization side, so it was probably inevitable we would take our different approaches to gaussian splats. I even waited impatiently through the spring until education site licenses got access to esri’s reality capture tools to see how they compared to some of the open source and trial software that I had moved toward.

Instead of going through things I have tried, I am going to highlight what I have used this month (mostly because otherwise the list is too long). My work falls into two use cases and one curiosity. The first use case is regular drone flights on most Sunday mornings over the various construction that is going on around campus. I have been doing these flights since I started my current job in summer of 2022, so I have 4 years of data sitting around that I am slowly referencing to get it into a time aware format for our facilities group to use with as-builts and other construction records. So, why not through some lovely gaussian splats into the mix. The morning we left for the UC I flew one of the current construction projects, processed the gaussian splat it in ArcGIS Reality for ArcGIS Pro and uploaded it to ArcGIS Online. The traditional geospatial workflow, but the problem is that if I want to share it in Superspl.at or if Sue wants to use it in PlayCanvas, we can’t. At the UC there was a lot of talk about the news item we discussed in the spring of industry support for glTF 2.0, but I still can’t export a ‘gaussian splat layer’ to a .ply or .sog (etc) as far as I can tell (please let me know if I am just missing this capability in my rush).

So, I am also generally running my traditional processing in Agisoft Metashape to get my ortho, but I can also export the aligned cameras as a COLMAP to take into a 3dgs training app. I have generally been using Lichtfeld Studio and I have had good results when I prepped well 🙂 I have also used RealityScan to align images and I have to admit that it is a great free way to process drone imagery even for traditional models in addition to a COLMAP export. I did also use PIX4DMatic for a trial and was happy with the results, but I am having trouble giving them money (no really, it was hard to contact an actual person. I had to do a virtual training session just so I could ask one of the trainers to get sales to contact me), so it isn’t in my daily tools at this point.

My second use case is working with 360 video, and this one I am still working on my process. I have video captures of some of the locations that Sue and I have been visiting over the last few years. The videos were collected for just that, video. So while I can use them for training 3dgs, they weren’t collected with masking me out in mind (I figured I would just generate an mp4 that focused anywhere but me after all). Some tools, especially online tools will ingest 360 video directly, even in proprietary formats like Insta360’s insv. But I have been using tools like pano2views to pull frames from the video and pull oriented photos from the 360 frame. Thanks again to Epic for making a tool that I can use for free. Once I have the extracted photos, I have been taking them into the various tools. But this has only been for quick tests since the campus data project is higher on my list of todos.

The final curiosity is using lidar point clouds with the imagery to develop 3dgs. I haven’t had the time to do this. I have, in the past, used this to create usdz files from meshes built on scans with iPad or iPhone Pro’s, but the app I used didn’t seem to maintain the raw imagery so I can’t reprocess it two years later, but I am still hoping (fruitlessly) that I am just missing a dataset somewhere. But, old projects in tools like Kiri Engine or PIX4DCatch can probably be able to be rerun to generate 3dgs from older collections though I haven’t had a chance to check this theory.

So, that is what I am using in July 2026. If you have a workflow or recommendations, let me know via email. I may have tried it in months before July, but I clearly didn’t stick with it this month at least. Oh…and I am currently trying out Postshot and aholo, but I just started with those (I know, I saved most people’s ‘go to’ for a later test).

The first splat below was created using Metashape to align and generate a COLMAP, then trained in Lichtfeld Studio and uploaded to SuperSpl.at

The splat below was generated in ArcGIS Reality for ArcGIS Pro and published to AGO.