When Gaussian Splatting Meets 19th Century 3D Images

Depending on which side of the English Channel / La Manche you sit on, photography was invented either by Englishman Henry Fox Talbot in 1835 or Frenchman Louis Daguerre in 1839.

By 1851, Englishman Sir David Brewster and Frenchman Jules Duboscq had perfected stereophotography. It led to an explosion of creativity in 3D photography, with the London Stereoscopic and Photographic Company becoming one of the most successful photographic companies of the era.

There are thousands of stereoscopic images hidden away in museum archives. For example, here's one from Birmingham, UK:

You probably don't have a stereoscope attached to your computer, but the 3D depth effect can be simulated by animating the two images.

Fast forward to 2023 and the invention of Gaussian Splatting. Essentially, using computers to work out 3D information when given multiple photos of a scene. It is magic - but relies on lots of photographs of a scene. Then, in 2024, Splatt3r was released. Give it two photos from an uncalibrated source, and it will attempt to reconstruct depth information from it.

Putting the above photo into the demo software gives us this rather remarkable 3D model as rendered by SuperSplat.

I think that's pretty impressive! Especially considering the low quality and low resolution of the images. How accurate is it? The pub is "The Old Crown" in Digbeth and is viewable on Google Streetview.

It's hard to get a perfect measurement - but I think that's pretty close.

Interactive Examples

Here's the image above.

Here are the Shoshone Falls, Idaho - from a series of photos taken in 1874.

This is Li Hung Chang from a stereograph taken in 1900.

Of course, it doesn't always produce great results. This is Roger Fenton's 1860 stereograph of the British Museum's Egyptian Room (Statue of Discobolus). Even with a high resolution photograph, the effect is sub-par. The depth works (although is exaggerated) but all the foreground details have been lost.

Background

Regular readers will know that I played with something similar back in 2012 - using similar software to recreate 3D scenes from Doctor Who. I also released some code in 2018 to do the same in Python.

Both of those techniques worked on screenshots from modern 3D video. The images are crisp and clear - perfect for automatically making 3D models. But neither of those approaches worked well with old photographs. There was just too much noise for simple code to grab onto.

These modern Gaussian Splatting techniques are incredible. They seem to excel at detecting objects even in the most degraded images.

Next Steps

At the moment, it is a slightly manual effort to pre-process these images. They need to be cropped or stretched to squares, artefacts and blemishes need to be corrected, and some manual tweaking of the final model is inevitable.

But I'd love to see an automated process to allow the bulk transformations of these images into beautiful 3D models. There are over 62,000 stereographs in the US Library of Congress alone - and no doubt thousands more in archives around the world.

You can download the images and models from my CodeBerg.