The first keynote today at #CVMP2025 is by Angela Dai @adai.bsky.social from 3DAI Lab at TU Munich

"Can Transformers speak geometry?"

There are lots of different 3D representations we use in learning but industry widely uses meshes.

𝗗𝗲𝗢𝗰𝗰-𝟭-𝘁𝗼-𝟯: 𝟯𝗗 𝗗𝗲-𝗢𝗰𝗰𝗹𝘂𝘀𝗶𝗼𝗻 𝗳𝗿𝗼𝗺 𝗮 𝗦𝗶𝗻𝗴𝗹𝗲 𝗜𝗺𝗮𝗴𝗲 𝘃𝗶𝗮 𝗦𝗲𝗹𝗳-𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝗠𝘂𝗹𝘁𝗶-𝗩𝗶𝗲𝘄 𝗗𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻
Yansong Qu, Shaohui Dai, Xinyang Li ... Rongrong Ji
arxiv.org/abs/2506.21544
Trending on www.scholar-inbox.com

🚀🚀🚀Announcing our $13M funding round to build the next generation of AI: 𝐒𝐩𝐚𝐭𝐢𝐚𝐥 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐌𝐨𝐝𝐞𝐥𝐬 that can generate entire 3D environments anchored in space & time. 🚀🚀🚀

Interested? Join our world-class team:
🌍 spaitial.ai

youtu.be/FiGX82RUz8U

Quoting one slides from @yann-lecun.bsky.social talk…

arxiv is filled by papers that treat symptoms (or not even!) without ever diagnosing the disease

My Vision Transformer lecture snippet

DAGM GCPR'25: 15 days.
AAAI'26 (abs): 67 days.
AAAI'26 (paper): 74 days.
3DV'26: 90 days.

This is a big/under-discussed deal given that most students are feeding assigned readings through LLMs instead of reading them in full.

Amid all the super-duper research advances, I am excited to share my video on a super-duper basic topic:

How to sample your signals?

Here we discuss sampling, why aliasing occurs, and how to prevent it. It's fascinating to view this through the lens of frequency analysis. youtu.be/fTJjPGaPsq4

I made a quick and dirty comparison between "optimal" and midpoint in a repo, feel free to try it out, and if you are a reprojcel, please let me know why the benchmark is biased, stupid, and wrong)!
github.com/Parskatt/tri...

When vision people do graphics, they call it "image synthesis." When graphics people do vision, they call it "inverse rendering". ;)

To read about, evaluate models, and use VSI-Bench, see:

Webpage: vision-x-nyu.github.io/thinking-in-...
ArXiv: arxiv.org/pdf/2412.14171
Eval Code: github.com/vision-x-nyu...
VSI-Bench: huggingface.co/datasets/nyu...
[n/n]

3D content creation with touch!

We exploit tactile sensing to enhance geometric details for text- and image-to-3D generation.

Check out our #NeurIPS2024 work on Tactile DreamFusion: Exploiting Tactile Sensing for 3D Generation: ruihangao.github.io/TactileDream...
1/3

Another gem from Bill Freeman, Katie Bouman & team 🌌

A differentiable rendering framework for direct #exoplanet imaging, leveraging wavefront sensing to refine starlight subtraction. Tested on JWST, it approaches noise limits and reveals faint planets like never before! 🚀

#ai #astronomy

I have just realised, that if LLM reads only the paper content, w/o meta-data, ALL the papers present results as novel and sota, so it is not easy to understand, which one is really new...

i recently gave a talk about questions like this at a workshop organized by the amazing "learning theory alliance" (let-all.com), and several students reached out to tell me how much my talk resonated with them.
link to the slides here:
docs.google.com/presentation...

I still don’t understand why it can be that distillation works, given the same data.

Is it a way to smuggle more computation into smaller model without looking at the data much more times?

How to design your presentation?

Presentation is an essential skill for academics.

After attending so many meetings, prelims, thesis defenses, research talks, and lectures, I’ve realized that I may have been approaching it all wrong ...

Some (bitter) lessons I learned. 👇

I recently gave a tutorial on the DUSt3R paper (web: dust3r.europe.naverlabs.com, paper: tinyurl.com/5t2ks575, code: github.com/naver/dust3r) in a research group meeting. In case you missed it, didn’t understand it or would like to hear some perspectives on why it’s such a cool idea, read on… 1/23