We round off #iccv25 with highlight paper Geo4D on Thur
TLDR: Geo4D repurposes a video diffusion model to reconstruct dynamic scenes in 4D
Paper: arxiv.org/abs/2504.07961
Project: geo4d.github.io
@chuanxiaz.bsky.social @dlarlus.bsky.social rlus.bsky.social @oxford-vgg.bsky.social gg.bsky.social
🧵9

Tuesday is Poster Session 2 with
HAMS3tR: Human-Aware Multi-view Stereo 3D Reconstruction @iccv.bsky.social
TLDR: HAMSt3R extends MASt3R to handle scenes involving people.
Paper: arxiv.org/abs/2508.16433
@weinzaepfelp.bsky.social @vincentleroy.bsky.social 🧵7/9

It's a marathon Sunday for @gabrielacsurka.bsky.social – co-author of PANst3R, Gabriela is also giving 2 invited wksp talks AND keynote dinner speech #WiCV 😅
CALIPOSE: 🏃https://sites.google.com/view/calipose2025/
CRoCoDL: 🏃‍♀️https://localizoo.com/workshop/
WICV: sites.google.com/view/wicv-ic... 🧵6/9

If you’ve never heard the story of how the #DUSt3R or *St3R family was born & what’s going on now - don’t miss @vincentleroy.bsky.social & his new talk at the #E2E3D workshop, also Oct 19th
Workshop E2E3D (afternoon): e2e3d.github.io 🧵5/9

This or some kind of PRoPe arxiv.org/abs/2507.10496

Yes exactly if you are dealing with unordered image collections. The same certainly cannot be said when dealing with videos where the displacement between two frames makes more sense.

Wrt to RoPE: People are applying it wrongly for multi-view, if you do it VGGT style you will mix the positions of all the images. You should do it independently (i.e. no rope between images).

And here is a link to the thesis itself: liu.diva-portal.org/smash/record...

Adding the fab view of Grenoble by night!