For the final plots showing PSNR vs training time, have you considered plotting on a log scale? It appears the main benefit is the reduced time required to reach a given PSNR. These close asymptotic tails seem to obscure this result. Also, using hours rather than seconds might be clearer 😅. Great work and Cool paper!
Im excited for a future where you can create a splat in real time with multiple cameras 😵💫 might even be cool for things like autonomous driving (or instead of security cameras, you mix them into a live splat and have security angels with vr headsets flying over crouds to detect anomalies)
Heck, military applications are also obvious here, but I don't like them :P might even be able to predict movement when out of line of sight? Like a deep 4D kalman filter @@ hoho. We are so fucked
I hope so, code please? :) I would love to attempt something with this as well, have an idea for an app/experience that will require me to better understand and be able to implement this tools. Thanks for the demonstration!