iiinova — where your digital self keeps evolving

Aditya Chetan @justachetan.bsky.social · 29/05/2026

If you are @cvprconference.bsky.social come find me at our poster: 🗓️ Sat, June 6 2026, 11:45 AM - 1:45 PM CDT 📍 ExHall F, Poster #222

010

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Project page 🔗: flat-pack-bench.github.io Huge thanks to my amazing collaborators: Eric Cai, Peeyush Kushwaha, Bharath Raj Nagoor Kani, Utkarsh Mall, @qianqianwang.bsky.social, @snavely.bsky.social, @bharathhariharan.bsky.social

000

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Finally, we test whether decomposition into simpler tasks (like humans) using agents with tracking and contact-reasoning VLM tools can help. The tools themselves struggle, and the agent, TVA, performs poorly, mostly abstaining and achieving only ~12% accuracy. 🤖(6/n)

100

Aditya Chetan @justachetan.bsky.social · 29/05/2026

We also find that models do not use videos effectively when answering questions. On an image-only version of the benchmark, humans perform worse as the questions need videos to answer, but model performance was preserved or improved! 🎥📸 (5/n)

100

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Probing Gemini 2.5 Pro’s thought traces revealed that it often loses object identity through camera motion, occlusion, and pose changes, or misreads when two parts actually make contact. The benchmark makes those failures visible. 👀 (4/n)

100

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Zero-shot results show that models lag far behind humans, who achieve 94.18% accuracy. The best model setting stays under 42%: InternVL3-78B reaches 41.03%, and GPT-5 reaches 37.71%, best amongst closed models. For context, chance performance is 26.74%. 📉(3/n)

100

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Flat-Pack Bench is a new video QA benchmark based on furniture assembly, with multiple-choice questions on internet videos, covering four tasks: temporal localization, ordering, mating, and tracking. We use frames from the video as visual prompts to point at exact parts. 🔍 (2/n)

100

Aditya Chetan @justachetan.bsky.social · 29/05/2026

Humans can watch tasks like cooking or assembly and reason about what happened, when, and between which parts. Can LVLMs do the same? We built Flat-Pack Bench to test this – and found there is still a long way to go. Accepted at #CVPR2026! 🎥🪑🧩(1/n)

131

Reposted by Aditya Chetan

Zhenjun Zhao @ericzzj.bsky.social · 27/05/2026

G3T Up! Gravity Aligned Coordinate Frames Simplify Pointmap Processing Bharath Raj Nagoor Kani, @snavely.bsky.social tl;dr: fine-tune VGGT on custom gravity-aligned 3D data->gravity-aligned pointmaps arxiv.org/abs/2605.27372

051

Reposted by Aditya Chetan

Tomer Ullman @tomerullman.bsky.social · 06/11/2025

It’s grad school application season, and I wanted to give some public advice. Caveats: -*-*-*-*  > These are my opinions, based on my experiences, they are not secret tricks or guarantees  > They are general guidelines, not meant to cover a host of idiosyncrasies and special cases

411761

Reposted by Aditya Chetan

Axel Bax @axel-bax.bsky.social · 28/08/2025

We are trying to create a list of in-copyright novels that contain maps. If you know of some, drop them in the thread below! 🧵👇

34146

Aditya Chetan @justachetan.bsky.social · 29/08/2025

Are fictional maps okay? If yes, the inheritance cycle by Christopher paolini, also the Throne of Glass series by Sarah J Maas

110

Reposted by Aditya Chetan

Bharath Hariharan @bharathhariharan.bsky.social · 12/06/2025

For those at CVPR, @justachetan.bsky.social will be presenting this poster tomorrow at 10:30 (Exhibit hall D, Poster #34). Come hear about why neural field derivatives are noisy, and how we resurrect image processing ideas for neural fields!

031

Aditya Chetan @justachetan.bsky.social · 10/06/2025

Thrilled to attend my first-ever #CVPR2025! 🎉 Please reach out if you would like to chat about neural fields, dynamic scenes, video understanding, or just generally about gaming, musicals, or ☕️ I will also be presenting our poster ⬇️ (Come visit!)

010

Aditya Chetan @justachetan.bsky.social · 10/06/2025

Happy to get feedback + questions! For more experiments and technical details, check out our paper! 😄

000

Aditya Chetan @justachetan.bsky.social · 10/06/2025

We also show improved performance in downstream applications like rendering, collision simulation, and PDE solving. (n/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

We show the effectiveness of our method in computing accurate normals and curvatures over a variety of challenging neural SDFs learned over the FamousShape dataset. Our approach shows a 4x improvement in gradients and mean curvature over the baselines. (6/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

Second, to enable smoother gradients directly with autodiff over the network, we propose a fine-tuning approach that can use any smooth gradient operator to smooth out the artifacts in the gradients. (5/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

To mitigate this noise, we propose a two-pronged solution. First, we leverage the classical technique of polynomial-fitting to fit low-order polynomials through the learned signal and take autodiff over the fitted polynomial. (4/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

What causes these artifacts? We note that signals learned by hybrid neural fields exhibit high-frequency noise (see FFT of a 1D slice of a 2D SDF), which gets amplified when we take derivatives using standard tools like autodiff. (3/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

Hybrid neural fields like Instant NGP have made training neural fields extremely efficient. However, we find that they fall short of being "faithful" representations, exhibiting noisy artifacts when we compute their spatial derivatives with autodiff. (2/n)

100

Aditya Chetan @justachetan.bsky.social · 10/06/2025

Check out our poster at #CVPR2025 on accurate differential operators for hybrid neural fields (like Instant NGP)! 🗓️ Fri, June 13, 10:30 AM–12:30 PM 📍 ExHall D, Poster #34 🔗 justachetan.github.io/hnf-derivati... 👉 cvpr.thecvf.com/virtual/2025... Details ⬇️ (1/n)

110

Reposted by Aditya Chetan

Brihi Joshi @brihi.bsky.social · 29/04/2025

Reasoning about the "why" behind user behavior can improve LLM personas! ✨🧠📈 📝Excited to share our new work: Improving LLM Personas via Rationalization with Psychological Scaffolds 🔗 arxiv.org/abs/2504.17993 🧵 (1/n)

1134

Reposted by Aditya Chetan

Anand Bhattad @anandbhattad.bsky.social · 29/03/2025

[1/10] Is scene understanding solved? Models today can label pixels and detect objects with high accuracy. But does that mean they truly understand scenes? Super excited to share our new paper and a new task in computer vision: Visual Jenga! 📄 arxiv.org/abs/2503.21770 🔗 visualjenga.github.io

75914

Reposted by Aditya Chetan

zhengqili.bsky.social @zhengqili.bsky.social · 06/12/2024

Introducing MegaSaM! Accurate, fast, & robust structure + camera estimation from casual monocular videos of dynamic scenes! MegaSaM outputs camera parameters and consistent video depth, scaling to long videos with unconstrained camera paths and complex scene dynamics!

16818