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👇

The French often mention the roundabout at Arc de Triomphe as one of the ultimate tests for a self-driving vehicle. I would say that the equivalent for humanoid robots would be to unwrap Chupa Chups Lollipop. Take that Embodied AI!

How to name your method: a comprehensive flow chart

I'm still puzzed by what type of memory would allow for such behavior: The girl disappeared for several seconds after full occlusion (sec. 13-14), however appeared once more at the end of the video. So it is clearly not LSTM style memory. Any ideas?

SCMs often assume overly-dense causal graphs in dynamic settings 👉⚽🥎, since any object interaction is a potential causal edge, making them hard to scale. In joint work with @turanorujlu.bsky.social @cgumbsch.bsky.social & Martin Butz we propose a new Causal Process Model to tackle this. Thread👇

Happy to organize #CoRL2025 Workshop! If you are interested in world models or learning simulators for robotics, consider joining! simulatingrobotworlds.github.io More info 👇

🚀 We’re pleased to announce our workshop at CoRL 2025: Learning to Simulate Robot Worlds! Workshop website: simulatingrobotworlds.github.io The workshop aims to cover topics like physics-grounded simulation, photorealistic digital twins, AI-controlled simulators, to learned neural world models.

🌍🤖 What is the best way to explore the world to learn a robust world model from high dimensional data? 🤖🌍 SENSEI learns to explore from humans by reusing semantics structure discovered by VLMs and exploring around most interesting for VLM states. #ICML2025

this post: ysymyth.github.io/The-Second-H...

🚀 Never miss a beat in science again! 📬 Scholar Inbox is your personal assistant for staying up to date with your literature. It includes: visual summaries, collections, search and a conference planner. Check out our white paper: arxiv.org/abs/2504.08385 #OpenScience #AI #RecommenderSystems

Interesting idea: RL with memory traces = exponential moving averages of latent encodings. arxiv.org/abs/2503.15200 O. Eberhard, M. Muehlebach, @claireve.bsky.social

Some hard evals. 1) IMC24 train -- Lizard. Reconstruction is good, but summer and winter are two separate nearby 2) IMC 24 train -- pond. Results are hard to tell. 3) Room with mirror - almost ideal Overall the results are very impressive, but SfM is not solved yet. 1/

It is pointless, but. My US friends - it will not stop after 4 years with next elections. Either you stop it yourself in some way, or there will be no fair elections next time. Democracy is fragile. P.S. If you like what is going on - enjoy.

u24.gov.ua Best what one can do as individual is it set a monthly target and donate it until the victory automatically. For example as standing order from your bank using SEPA: koloua.com/en/donate#sepa

By popular demand, we are extending #ICCV2025 coverage to Bluesky. Stay tuned!

Re-posting is appreciated: We have a fully funded PhD position in CMC lab @cmc-lab.bsky.social (at @tudresden_de). You can use forms.gle/qiAv5NZ871kv... to send your application and find more information. Deadline is April 30. Find more about CMC lab: cmclab.org and email me if you have questions.

Our team @GoogleAI is hiring an intern. We are interested in having LMs understand and respond to users better. Topics include: teaching LMs to build “mental models” of users; improving LM's reasoning capability over long contexts. @GoogleAI internship deadline is Feb 28.

computer vision researchers’ conversations with friends be like

Who would be "Japan" to have an example?

Exciting news! MegaSAM code is out🔥 & the updated Shape of Motion results with MegaSAM are really impressive! A year ago I didn't think we could make any progress on these videos: shape-of-motion.github.io/results.html Huge congrats to everyone involved and the community 🎉

Elon Musk pusht die Pro-Kreml Parteien AfD und BSW im For You Feed von Twitter. Sie tauchen öfter im For You Feed auf als ihrem Anteil entspricht. Die Forscher haben dafür Test-Accounts erstellt, die allen Politikern auf X folgen. BSW und AfD waren krass überrepräsentiert.

European Parliament member 👇🏽

🚀 We’re hiring! Join Bernhard Schölkopf & me at @ellisinsttue.bsky.social to push the frontier of #AI in education! We’re building cutting-edge, open-source AI tutoring models for high-quality, adaptive learning for all pupils with support from the Hector Foundation. 👉 forms.gle/sxvXbJhZSccr...

I'm curious to which category "making things with LLM/code-gen" falls. I would guess that it still feels that the Python script entirely generated LLM (e.g., R1) still feels satisfying as you can modify any aspect of it, even if you don't have to most of the time for simple tasks.

✨ The VIS Lab at the #University of #Amsterdam is proud and excited to announce it has #TWELVE papers 🚀 accepted for the leading #AI-#makers conference on representation learning ( #ICLR2025 ) in Singapore 🇸🇬. 1/n 👇👇👇 @ellisamsterdam.bsky.social

Also as I may have VISA problems once more I want to apply as early as possible, do you know whom I should contact for an early invitation letter?

(1/3) Happy to share LUDVIG: Learning-free Uplifting of 2D Visual features to Gaussian Splatting scenes, that uplifts visual features from models such as DINOv2 (left) & CLIP (mid) to 3DGS scenes. Joint work w. @dlarlus.bsky.social @jmairal.bsky.social Webpage & code: juliettemarrie.github.io/ludvig

Asking for a friend: how many hours it is OK to spend trying to fit answers including references to 1-page CVPR rebuttal answer and do not feel like I'm spending public money for something really useless?

Receiving #ICLR2025 decisions and then #CVPR2025 reviews shortly after

Useful thread to remind ourselves that while evidence can look like supporting opposite strong priors, in reality, evidence can suggest that both priors are incorrect and more evidence needed. P.S. 3 figures from the thread could be a nice triptych that would be a good reminder to stay open-minded.

I have switched from X and Facebook to Bluesky. I will delete both my X and Facebook accounts in a little bit.

After reading about Genesis, I think quite a few features seem to come directly from Taichi ("parallel programming language for high-performance numerical computation") that uses pytorch backend. With all the marketing here, I'm most puzzled I haven't heard about Taichi before

100%

Neural Assets poster is happening now. Join us at East Exhibit Hall A-C #1507

Can we enhance the performance of T2I models without any fine-tuning? We show that with our ReNO, Reward-based Noise Optimization, one-step models consistently surpass the performance of all current open-source Text-to-Image models within the computational budget of 20-50 sec! #NeurIPS2024

Today I’ll be presenting the poster #5110 from Dingling Yao applying CRL to climate and dynamical systems identification “Marrying Causal Representation Learning with Dynamical Systems for Science” West Ballroom A-D, starting at 16:30 Come check it out!

My earlier post was not intended to dunk on the Sora model or on Jim Fan, but to call out, and advise against, the ML community's frequent vice of resorting to "wishful mnemonics" – such as calling a video generation model a "physics engine". x.com/chrisoffner3...

Exciting opportunity to intern with me 🙋‍♀️ & the entire research unit lead by @jhartford.bsky.social 🧠 on single-cell representation challenges 🦠 in gene perturbation imaging data 🧬 Come join Valence Labs London office 🇬🇧 powered by @recursionpharma.bsky.social: job-boards.greenhouse.io/valencelabs/...

LoRA et al. enable personalised model generation and serving, which is crucial as finetuned models still outperform general ones in many tasks. However, serving a base model with many LoRAs is very inefficient! Now, there's a better way: enter Prompt Generation Networks, presented today #BMVC

go.bsky.app/5KXsVXb Created Starter Pack for VISLab, if you see more people whom I should add suggest here

Introducing Generative Omnimatte: A method for decomposing a video into complete layers, including objects and their associated effects (e.g., shadows, reflections). It enables a wide range of cool applications, such as video stylization, compositions, moment retiming, and object removal.

Am I the only one who is thinking that one list of followers is an outdated way to decide what should be in the timeline? Why not multiple lists?

I think such location+topic based lists are really useful! Found a lot of people with whom I was previously connected!

Turns out aria-glasses are a very useful tool to demonstrate actions to robots: Based on egocentric video we track dynamic changes in a scene graph and use the representation to replay or plan interactions for robots 🔗 behretj.github.io/LostAndFound/ 📄 arxiv.org/abs/2411.19162 📺 youtu.be/xxMsaBSeMXo

Fully agree with this part. Always like to think of "upper bounds" with privileged info and do test them - quite often the results are not as expected and then one starts to understand the problem better. While this could seem slower, in the long run it often leads to more interesting results

Looking forward to seeing what is possible to build on top of such "particle" representations. While conceptually simple, they are one step closer to represent scenes (underlying causal structure) not videos (mixture of the many factors together), and could be useful for robotics tasks.

I would say it could be the opposite. While having a "small" agent, it should have the capacity to construct compositional sub-models of the environment on the fly.