New feature extractor just dropped

I am hiring a PhD & postdoc to work together with me at KTH on probabilistic machine learning. Both positions are fully funded and part of WASP. I will be attending @euripsconf.bsky.social, if you are around and want to talk about the positions or what we do at KTH, then ping me and we can meet.

Getting journal rejections like

SAM 3D: 3Dfy Anything in Images SAM 3D Team et al? tl;dr: in title. 8-stage training, dataset, human labeling. Do not read tl;dr, read whole paper arxiv.org/abs/2511.16624

RoMa v2 is now out! (github.com/Parskatt/rom..., arxiv.org/abs/2511.15706) Here are the main improvements we made since RoMa:

They trained a vanilla Vision Transformer (ViT), to perform image-to-image mapping, solely on Abstraction and Reasoning Corpus (ARC) data and generalizes to unseen tasks through test-time training. The model achieves 60.4% accuracy on the ARC-1 benchmark, outperforming existing methods.

It's now recommended by ICLR to flag poor/LLM reviews to ACs xcancel.com/iclr_conf/st... Quite interesting how this will turn out given that pangram estimates 20% of all reviews to be LLM generated iclr.pangram.com/reviews

Criticising a paper without concrete suggestions for how to attain inner calm is bad reviewing tbh.

Last week we launched IMC2025-Ongoing on @kaggle.com The dataset is exactly as in IMC2025, but the competition is on-going for a year, making it better for academic leaderboard and persistency. kaggle.com/competitions... 1/2

What is the "author registration" for CVPR 2026? Got no hits from a ctrl-f on the Author Guidelines page @cvprconference.bsky.social

blog.arxiv.org/2025/10/31/a... FYI the blog post for the updated policy is out. Our llm future is dire:/

Want to generate point clouds? Our (w/ Ernst Röell) #NeurIPS2025 paper shows a topology-driven method that is fast, simple, and high-quality. 🔥 1/n

There's an increasingly popular argument saying that for 3D point clouds with rotational ambiguity (e.g. molecular data) standard transformers are preferable over equivariant models for compute reasons. With “Platonic Transformers” we show why this doesn't have to be true. arxiv.org/abs/2510.03511

Proud to be part of work on “Certifiably Optimal Anisotropic Rotation Averaging” that is an #ICCV2025 award candidate. If you're at the conference, come and learn about it — Carl Olsson gives a presentation today at 1:30pm (Oral 4B) in Kalakaua Ballroom; poster #305 is in Exhibit Hall from 2:45pm.

Lund University is hiring a Postdoctoral Fellow in Earth Observation! 🌍🛰️ Postdoc on vegetation fires and land abandonment using remote sensing and spatio-temporal analysis. 3-year, full-time research position. Apply by November 17, 2025. climateainordics.com/job-openings...

Turns out NLP is just vision

Poor eigenvalues can't make it from 1 to -1. :(

Do you trust RANSAC stopping criterion? - Yes, confidence=0.99 FTW - No, max_iter FTW - WTF are you talking about? Answer in comments.

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.

scipy deprecates `sph_harm` and replaces it with `sph_harm_y` where `n` and `m` have switched order and `theta` and `phi` have switched meaning 🙃

Images (or image patches) are secretly multi-channel signals over groups. Below, the dihedral group of order 8: reflecting/rotating the image permutes the values in the magenta vector. So we can reshape the image into 8-tuples that all permute according to the dihedral group (edge case diagonals).

Nice LLM generated citation found by @davnords.bsky.social. I wonder who M. Lindberg and A. Andersson are...

The opportunities and risks of the entry of LLMs into mathematical research in one screenshot. I think it is clear that LLMs will make trained researchers more effective. But they will also lead to a flood of bad/wrong papers, and I'm not sure we have the tools to deal with this.

A simple argument for equivariance at scale: 1) At scale, token-wise linear layers dominate compute. 2) Token-wise linear equivariant layers implemented in the Fourier domain are block-diagonal and hence fast.

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...

Going to recycle an extension of this thread, but to uncertainty principles: The idea here was that almost all the listed objects arise from choosing a specific operator encoding a structure of interest (a symmetry, a statistic, a geometry, or dynamics) and then seeking its eigenfunctions. [+]

Badness is an underrated word for loss functions

These funny layernorm (and layerscale) weight values are present in early DINOv3 layers as they were in DINOv2 by the way :)

Have you ever asked yourself how much your favorite vision model knows about image capture parameters (e.g., the amount of JPEG compression, the camera model, etc.)? Furthermore, could these parameters influence its semantic recognition abilities?

Introducing DINOv3 🦕🦕🦕 A SotA-enabling vision foundation model, trained with pure self-supervised learning (SSL) at scale. High quality dense features, combining unprecedented semantic and geometric scene understanding. Three reasons why this matters👇

Academic Articles Assessed In-silico (AAAI).

EurIPS is coming! 📣 Mark your calendar for Dec. 2-7, 2025 in Copenhagen 📅 EurIPS is a community-organized conference where you can present accepted NeurIPS 2025 papers, endorsed by @neuripsconf.bsky.social and @nordicair.bsky.social and is co-developed by @ellis.eu eurips.cc

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'm @ ICML, see you there :)

> (3189 additional authors not shown) ...jfc google.

The sycophantic tone of ChatGPT always sounded familiar, and then I recognized where I'd heard it before: author response letters to reviewer comments. "You're exactly right, that's a great point!" "Thank you so much for this insight!" Also how it always agrees even when it contradicts itself.

We just released COLMAP v3.12, which adds long-awaited, end-to-end support for multi-camera rigs and 360° panoramas 👀 COLMAP just got better at handling your robotics, AR/VR, or 360 data - try it yourself and let us know! github.com/colmap/colma... Kudos to Johannes & team for this great work 🚀

Dense Match Summarization for Faster Two-view Estimation Jonathan Astermark, Anders Heyden, Viktor Larsson tl;dr: use clustering to reduce RANSAC time when using dense methods like RoMa. Kudos for eval on WxBS. P.S. now the same, but for BA? arxiv.org/abs/2506.028...

UFM: A Simple Path towards Unified Dense Correspondence with Flow Yuchen Zhang et 11 al tl;dr: wide-baseline and optical flow data help each other. DINO layer selection. No matching, just flow regression with RoMA- loss. A LOT of ablations arxiv.org/abs/2506.09278 1/

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...

Reminder that the MSCA postdoctoral program exists. If you have a PhD and want to work in a European lab, you have until September to apply. Just contact them now. ec.europa.eu/info/funding...

Our recent paper on universality classes of equivariant networks arxiv.org/abs/2506.02293 enjoyed being part of it (and more to come!).

The war on science in the US is already having an effect on private sector research like AlphaFold. Bears repeating but the private sector builds on top of things created by academic research for the public good. This hurts everyone.

We're having a fun time with equivariance and modern vision architectures.

"Can it be read in the morning?"

Apparently AI in Danish is "kunstig intelligens", which sounds like "weird intelligence" in Swedish ("konstig intelligens").

What is mathematical beauty? tug.org/books/review...

I got the best teaching evals of my career for a large course last semester— along with three teaching awards— after rebuilding my intro class around in-class handwritten essays. I simply do not believe that this is unsolvable or that students don't care about actually learning to do their work.