I wonder if some move is going on from trying NeRF to the new tech of 3D Gausian Splatting?
My observation on Mapillary is that NeRF did not really pick-up.
I happened to come across a YouTube video recently about using a DJI Neo for such which caught my interest. Kiri Engine provides photogrammetry, NeRF and 3DGS. I am experimenting with it using my 2010 through 2014 imagery for Microsoft Research’s Photosynth. These didn’t NeRF well with Mapillary. First 3DGS results are not too bad.
Kiri Engine recently got API functionality and lists Meta as a client. I wonder what is going on?
If Mapillary continues on 3D reconstruction, then I suggest focusing the effort on objects instead of spaces. Having statues and like points of interest on the map in 3D is a great thing, I think. A lot of that imagery is likely available already.
@JoscTr - agreed that this field is moving quickly and that Gausian Splatting seems like promising next generation technology - this is something we’ll look at - thank you!
@boris - appreciate your response. If I understood and remember correctly, Kiri shares a lot on Github. Maybe it is of use in developing Meta’s own processing.
I should have posted this earlier on. Last year I have been experimenting a bit with 3DGS, using old 360 street level video captures. I have used DJI Studio to reframe the 360 videos and Kiri web to create the 3DGS’. I have used Supersplat to clean them up manually. See the results on my Supersplat Gallery profile joscelin-ost360vr - SuperSplat
Thank you, Boris. I feel that if a dedicated pipeline would be built, maybe even with a dedicated mobile app, plenty of people would join the crowd sourcing of points of interest.