An example of how DUSt3R can do "impossible matching": given two images without any shared visual content (my office, obviously never seen at training), it can output an accurate reconstruction (no intrinsics, no poses!) in seconds

For reference, here is how my office truly looks like

DUSt3R blog is also available

I thought you must be joking so I ran it on these two of my own pictures (excuse my dirty dishes 😉) and wow, it works beautifully.

I gave it a try, and the result was amazing! left is our 3d-printed #fnirs headcap reconstructed from 10 photos using #dust3r - took 1 min on a 4070 on my laptop right is my attempt yesterday using #visualsfm with 30 photos after dense recon sfm has gone a long way! well done!

Thanks 🙏 but note DUST3R has never seen human bodies during its entire training. Since Dust3r is a very data-driven approach by nature, reconstruction results on human bodies can be relatively sloppy and disappointing.

Can you explain how this is working in this case? Seems too good to be true

The answer is simple: witchcraft🪄🧙‍♀️

@chriswolfvision I love it! Impossible matching in trained feature space! Beautiful work. 👏

@chriswolfvision Thanks 😀 Kuddos to @riverakid1 @Vinc3nt_Leroy Yohann Cabon, and @bchidlovskii

This looks incredible! Can you give any estimates on inference runtimes for an image pair? I couldn't find any numbers in the paper.

something like 20ms on a A100 GPU?