These gentlemen show how not only colmap but also VGGT fail on spherical motion in their ICCV oral paper "Uncalibrated Structure from Motion on a Sphere".

I wonder if it is just a data issue for VGGT or if it is deeper than that. I mean VGGT was trained on mostly synthetic data. What do you think?

Obligatory picture with the poster :)

Presented with my co-supervisor @bokmangeorg.bsky.social at #ICML25, fun time!

Perhaps you'd be interested in our work on an adjacent topic, instead of optimizing the shape of ViTs we propose a sparse structure of the linear layers induced by the inductive bias of octic equivariance, significantly reducing FLOPs. Shameless self-plug :) (arxiv.org/abs/2505.15441)

Btw @bokmangeorg.bsky.social made some illuminating visualizations to show how our PatchEmbed leads to features that transform predictably under octic actions.

In particular, we achieve c. 40% reduction in FLOPs and increased throughput while improving performance for both supervised and unsupervised vision tasks.

We do so by constructing ViTs that use these computational savings early in the network and use regular blocks in the latter half.

We run a kernel constrained PatchEmbed to map an image to steerable ViT features and make use of the block-diagonalization of linear layers induced by the isotypical decomposition to achieve computational gains.

Want stronger Vision Transformers? Use octic-equivariant layers (arxiv.org/abs/2505.15441).

TLDR; We extend @bokmangeorg.bsky.social's reflection-equivariant ViTs to the (octic) group of 90-degree rotations and reflections and... it just works... (DINOv2+DeiT)

Code: github.com/davnords/octic-vits

GPT 4o's new image capabilities seem to be liked. The insinuation from OpenAI seems to be that it is not based on diffusion. I wonder how their work relates to the infamous NeurIPS paper "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction" (arxiv.org/abs/2404.02905)