We are introducing MuM, a feature encoder (ViT-L) tailored for 3D vision tasks. TLDR; Spiritual successor to CroCo with a simpler multi-view objective and larger scale. Beats DINOv3 and CroCo v2 in RoMa, feedforward reconstruction, and rel. pose. arxiv.org/abs/2511.17309 github.com/davnords/mum

Johan out here carrying Swedish CV academia with a 10 author absolute masterclass. Only the avatar can master all four elements (LiU, CTH, LTH and UvA). Very strong and versatile model overall. RoMa v2 can match anything you'd like. Even @bokmangeorg.bsky.social 's hand annotated satellite images.

RoMa v2: Harder Better Faster Denser Feature Matching @parskatt.bsky.social et 11 al. tl;dr: in title. Predict covariance per-pixel, more datasets, use DINOv3, adjust architecture. arxiv.org/abs/2511.15706

Turns out NLP is just vision

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?

Pro tip: For good Halloween vibes, use non-normalized RoPE on images larger than your training resolution and larger than the composite period of some of the RoPE-rotations. You might get scary ghost structures in your features.

RoMa now on PyPI under name of `romatch`

Towards the Next Generation of 3D Reconstruction @parskatt.bsky.social PhD Thesis. tl;dr: would be useful in teaching image matching - nice explanations. (too) Fancy and stylish notation. Cool Ack section and cover image. liu.diva-portal.org/smash/record...

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

How to name your method: a comprehensive flow chart

Born too late to explore the earth. Born too early to explore the galaxy. Born just in time to \nabla_{\theta}f_{\theta}

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

Tomorrow at ICML [Tuesday 4:30 pm, poster W-213] @davnords.bsky.social and I will present our spotlighted flop paper. Come by and let us try to convince you that equivariant nets should be standard in vision tasks due to computational benefits! bsky.app/profile/bokm...

I am at ICML, happy to connect with all of you :)

And today I'm presenting this work at #CVPR2025! 🗓️ Date: 16:00-18:00, Fri, Jun 13 (Today) 📍Place: Poster #115 in Session 2 (ExHall D) 💻 Code: github.com/ericssonrese...

As models become larger, more of their compute is spent in the MLP. Turns out that this is perfect for octic equivariance, as our biggest gains are there!

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)

New paper! We merge SfM reconstructions with point cloud registration. Link: arxiv.org/abs/2503.17093 Code: Not yet public, but coming later.

New paper! (arxiv.org/abs/2503.13433), we look into improving the threshold roubustness of Random Sample Consensus (RANSAC) through (less biased) inlier noise scale estimation.

Introducing VGGT (CVPR'25), a feedforward Transformer that directly infers all key 3D attributes from one, a few, or hundreds of images, in seconds! Project Page: vgg-t.github.io Code & Weights: github.com/facebookrese...

Introducing DaD, Part 2, a pretty cool keypoint detector.

We made a new keypoint detector named DaD, paper isn't up yet, but code and weights are: github.com/Parskatt/dad

Common beliefs about equivariant networks for image input include 1) They are slow. 2) They don’t scale to ImageNet. 3) They are complicated. In my opinion, these three are all false. To argue against them, we made minimal modifications to popular vision models, turning them mirror-equivariant.