Not an explicit field expert, but pretty well into computer graphics(games) and been reading a bunch of papers in the field over the years.
Classic photogrammetry was always a mixed bag in terms of results (especially if trying to construct meshes), but even before NeRFs (Neutral Radiance Fields) and Gaussian Splatting there was a ton of work using neural nets to handle various parts and I doubt that many modern tools avoid using them.
So in a way, these fields actually made use of neutral nets/"AI" (honestly more relevant imho than most of the LLM stuff).
I agree having worked directly with it (long personal project). Photogrammetry is so error prone (get a sign wrong in your sensor fusion/SLAM code and you're cooked for 2 weeks) and I got into it right in the last years before models that could compete were introduced (reactions from PhDs who spent years working on photogrammetry were not nice as you can imagine). Once the models came out they were set to take over, I remember being blown away from the quality of depth maps alone.
I think putting AI front and center in the marketing like this is a public relations move by Microsoft to brush up the image of AI in the general public.
"Microsoft provided the AI tech needed to process and analyze Iconem’s vast amount of photogrammetry data used to create the digital twin of St. Peter’s Basilica. Microsoft’s AI for Good Lab contributed advanced tools that refined the digital twin with millimeter-level accuracy, and used AI to help detect and map structural vulnerabilities like cracks and missing mosaic tiles."
Photogrammetry has rebranded itself to "Spatial AI".
Not an explicit field expert, but pretty well into computer graphics(games) and been reading a bunch of papers in the field over the years.
Classic photogrammetry was always a mixed bag in terms of results (especially if trying to construct meshes), but even before NeRFs (Neutral Radiance Fields) and Gaussian Splatting there was a ton of work using neural nets to handle various parts and I doubt that many modern tools avoid using them.
So in a way, these fields actually made use of neutral nets/"AI" (honestly more relevant imho than most of the LLM stuff).
True, although most of the snazzy NeRF and Gaussian Splatting papers still rely on good old COLMAP on the backend lol
I agree having worked directly with it (long personal project). Photogrammetry is so error prone (get a sign wrong in your sensor fusion/SLAM code and you're cooked for 2 weeks) and I got into it right in the last years before models that could compete were introduced (reactions from PhDs who spent years working on photogrammetry were not nice as you can imagine). Once the models came out they were set to take over, I remember being blown away from the quality of depth maps alone.
I think putting AI front and center in the marketing like this is a public relations move by Microsoft to brush up the image of AI in the general public.
"Microsoft provided the AI tech needed to process and analyze Iconem’s vast amount of photogrammetry data used to create the digital twin of St. Peter’s Basilica. Microsoft’s AI for Good Lab contributed advanced tools that refined the digital twin with millimeter-level accuracy, and used AI to help detect and map structural vulnerabilities like cracks and missing mosaic tiles."