Noah Snavely

That said, I don't actually mind "the supplemental". It sounds okay to my ear!

I suppose "supplement" would be okay, and shorter. Or "appendix" might sound nicer.

Page 2

I found this old bassoon sheet music in storage, and tried to have Gemini scan it to PDF. I think it is getting better with music? But still goes off the rails. (ChatGPT does even worse.)

But being metrics-driven seems a bit dissonant with the goals of the community, like, I don't think we should feel like "our business is generating reviews". (This is separate from the question of whether IEEE should be the arbiter of these things.)

Very interesting! That accords with the vibe I've gotten from recent conference town halls, where reviewing seems very numbers driven (success = 100% on time reviews), as if we were in a corporate boardroom celebrating our balance sheet. That's not a knock on organizers, who have a difficult job.

(Maybe more concretely, I'll propose that it shouldn't be seen as the end of the world and some shocking disaster if some papers have two reviews.)

Though you are totally right that according to this model, the break-even point could be much higher than 3 (though you also have to factor in the increased work, somehow).

(I also feel like we're way past the point where peer review is really proceeding in a super-scientific way at the big conferences, and we should accept that and try to take things a bit less seriously, though without throwing up our hands altogether.)

(My secret motivation is just that I feel I'm always reading invectives against reviewers, and the consequences against negligence are ratcheting up, e.g., desk rejection. I want to send a big love note to reviewers as an AC this year and am looking for ways to decrease pressure.)

And I don't know if the decisions are necessarily better as a result of the insistence on 3 reviews per paper.

Good point! But there's also a cost to the community as a function of the number of reviews. More work, more angst. I sense a lot of anger against reviewers, and a lot of pressure (including time pressure), which may just lead to LLM use.

Interesting! I was imagining that something like this could be the case: as the total number of reviews goes up, the average quality goes down. If so, it's not clear to me that 3 is the sweet spot.

I was thinking it may be a neat experiment to require fewer reviews and see if something breaks.

The CVPR process seems driven by the goal of extracting 3 reviews for each paper, a goal that seems to lead to a lot of angst. Is there a reason why 3 is a magic number? Why not two, or even one (assuming the quality is high)?

If you are looking for a new podcast to get into, I recommend Topics with Michael Ian Black and Michael Showalter. Even though they are comedians, they aren't afraid to dive into some weighty topics! www.earwolf.com/show_archive...

Thank you!

Thank you!

Any of the above. Are there computer vision papers that have beautiful figures whose intention is to explain how a system works? (E.g., network architecture, system diagram, etc). High-level organization, as well as low-level style (should I use rounded boxes? What colors should I use?) included

What are some examples of computer vision papers that have attractive system diagrams?

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis.

Today I want to share two new works on this topic:

Eliciting higher alignment: arxiv.org/abs/2510.02425
Unpaired learning of unified reps: arxiv.org/abs/2510.08492

1/9

#TTT3R: 3D Reconstruction as Test-Time Training
TTT3R offers a simple state update rule to enhance length generalization for #CUT3R — No fine-tuning required!
🔗Page: rover-xingyu.github.io/TTT3R
We rebuilt @taylorswift13’s "22" live at the 2013 Billboard Music Awards - in 3D!

We present a new approach to inference-time scene optimization, which we name Radiant Triangle Soup (RTS) www.arxiv.org/abs/2505.23642. Also check out really great concurrent work from Held et al. @janheld.bsky.social, Triangle Splatting arxiv.org/abs/2505.19175

🧠How “old” is your model?

Put it to the test with the KiVA Challenge: a new benchmark for abstract visual reasoning, grounded in real developmental data from children and adults.

🏆 Prizes:
🥇$1K to the top model
🥈🥉$500
📅 Deadline: 10/7/25
🔗 kiva-challenge.github.io
@iccv.bsky.social

(ChatGPT claims that this piece is Twinkle Twinkle Little Star, while Gemini says it is Do-Re-Me.)

ChatGPT and Gemini both seem to struggle with sheet music. They both insist that this excerpt is in D major (2 sharps), and resist any attempt to tell them that there 3 sharps in the key signature. I think this is really cool and interesting!

Think LMMs can reason like a 3-year-old?

Think again!

Our Kid-Inspired Visual Analogies benchmark reveals where young children still win: ey242.github.io/kiva.github....

Catch our #ICLR2025 poster today to see where models still fall short!

Thurs. April 24
3-5:30 pm
Halls 3 + 2B #312

Dynamic Camera Poses and Where to Find Them

Chris Rockwell, @jtung.bsky.social, Tsung-Yi Lin, Ming-Yu Liu, David F. Fouhey, Chen-Hsuan Lin

tl;dr: a large-scale dataset of dynamic Internet videos annotated with camera poses

arxiv.org/abs/2504.17788

1/6 🔍➡️ How to transform standard videos into immersive 360° panoramas? We've designed a new AI system for video-to-360° panorama generation!

Our key insight: large-scale data is crucial for robust panoramic synthesis across diverse scenes.

We have released the Stereo4D dataset! Explore the real-world dynamic 3D tracks: github.com/Stereo4d/ste...

This is really nice work on visual discovery from @boyangdeng.bsky.social!