Kocur, Aung, and Flood, et al., "Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth" You can use camera poses to benchmark depth estimators as well. Each model seem ot have their own bias towards which kind of scene they are best at.

VGGT-Omega paper is so pleasant to read. Unlike most of papers which I just skim through, I enjoy this one and read every sentence. @jianyuanwang.bsky.social et al - great job!

The more I experiment, the more I believe that "mid layers know more geometry" etc is mostly an artifact of probing frozen models. Same about the pooling. GeM is great for frozen backbone. If you can train, use attention.

📢📢📢 While you are all engaging in lively discussions during the rebuttal phase, we want to announce that the #ECCV2026 conference dates have been updated! It has been decided that the conference will end on Saturday the 12th September. Please have this in mind when making travel plans

Apparently #CVPR2026 had an outstanding AC initiative. Well done, @davidpicard.eurosky.social!

We are implementing a similar policy at @arxiv.bsky.social. If there is incontrovertible evidence of LLM slop in a paper, this means the authors did not take the time to read the LLM output and we can't trust anything else in the paper. Penalty is 1 year ban from arXiv followed by...

Claude: "Where I'd put my money" Me: "Bro, you don't have any money"

The quantification trap, or how well-intentioned attempts at intersubjectivity breed institutionalized postmodern optimization. (Part 2 in my series on what I wish I had more clearly articulated in The Irrational Decision.)

Claude, after his clone was working on a feature for an hour: "The branch is now 3 weeks behind on a pipeline-level feature"

How dare you exploit academic labor; that’s our job!

Agent: Let’s start slow and keep it simple. Me: Great. Now make it robust. Agent: Absolutely. I’ll preserve backward compatibility with all the mistakes we made 30 seconds ago.

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

NeurIPS'26 (abs): DL today, good luck (2h)! BMVC'26 (abs): 16 days. CORL'26 (abs): 18 days. CORL'26 (paper): 21 days. BMVC'26 (paper): 23 days.

Yes.

Match-Any-Events: Zero-Shot Motion-Robust Feature Matching Across Wide Baselines for Event Cameras Ruijun Zhang, Hang Su, Kostas Daniilidis, Ziyun Wang tl;dr: spatio-temporal transformer aggregation -> (ViT+DPT) for matching. New Event-MegaDepth dataset arxiv.org/abs/2604.18744

Triangulation of Points Constrained to a Plane @petrhruby.bsky.social Elima Shehu tl;dr: use DAv2 to get planar regions -> apply new solvers for those to get more precise triangulation. arxiv.org/abs/2604.27246

If you are going to use LLM to do the paper review (don’t), at least do it properly: - use Deep research mode - ask Claude/ChatGPT write a detailed deep research prompt first.

OK let’s see if people are better than Claude in literature review. Is there a paper, comparing output representation vs supervision? E.g. there is a consensus that predicting rays is better, than K matrix directly. But what is better- supervising K with rays, or rays with K?

Non-Minimal Sampling and Consensus for Prohibitively Large Datasets Seong Hun Lee, Patrick Vandewalle, Javier Civera tl;dr: you split (huuuge) tentatives list, and run RANSAC of each. How do you combine winning hypothesis? Truncated Lp cost seems the best. arxiv.org/abs/2604.22518

ArchSym: Detecting 3D-Grounded Architectural Symmetries in the Wild Hanyu Chen, @ruojin.bsky.social Steve Marschner, @snavely.bsky.social tl;dr: dataset + finetuned VGGT for symmetry detection arxiv.org/abs/2604.22202

Long-tail Internet photo reconstruction Yuan Li, Yuanbo Xiangli, Hadar Averbuch-Elor, @snavely.bsky.social @ruojin.bsky.social tl;dr: Cleaned-up version of MegaScenes + MoGe2-based depth refinement. +sparse sampler for emulate hard scenes. arxiv.org/abs/2604.22714

Li et al., "Long-Tail Internet Photo Reconstruction" MegaDepth-X, a new curated dataset for in-the-wild reconstruction that enhances long-tailed distributions. Many feed-forward estimators are colmap distillations, and this one makes the data prep more modern, which matters.

📣 ELLIS | ELLIOT Summer School on Autonomous Driving 🚗 Super happy to announce the hottest event this summer. Apply now! 📅 September 15-18, 2026 📍Barcelona, Spain 🌐 pages.cvc.uab.es/ssadriving/ @ellis.eu, ELLIS-Barcelona, CVC-Barcelona, @valeoai.bsky.social ai.bsky.social, @elliot-eu.bsky.social

ELViS: Efficient Visual Similarity from Local Descriptors that Generalizes Across Domains @psuma.bsky.social @gkordo.bsky.social @skamalas.bsky.social @gtolias.bsky.social tl;dr: SuperGlue, but single score on light-weight top descriptors. Most important thing is dustbin. arxiv.org/abs/2603.28603

Unlocking the Power of Critical Factors for 3D Visual Geometry Estimation Guangkai Xu et 6 al. tl;dr: VGGT finetuning ablation. 1. More data good, even if noisy 2. Confidence and gradient losses bad. 3. tying pcl, camera, intrinsics -- good . 4 high resolution trick. arxiv.org/abs/2604.21713

Gemini combination UX of hitting rate limits + deep research mode is the extreme stupidity. Recipe for the failure: You: prompt Gemini: Sorry, rate limit, try again as 12:00 You, at 12:00: Now? Gemini: Ignores the first prompt and does a deep research on "now" aka news digest?

Poorly tuned or surprisingly hard to beat? Bad luck that I would never figure out unless they a (Gaussian) filter one

🇧🇷 Presenting our ICLR 2026 paper “Efficient Probing” (EP) today! ❓What if linear probing is asking the wrong question? 🥳 EP is a lightweight attention probing method that better evaluates local, patch-level representations from models like MIM. 📍Friday 24 April, P4-#3713, 15:15–17:45

"The one remaining piece of stupidity" (c) Claud, after reviewing codebase.

🚨 Efficient Local Visual Similarity (ELViS) @ #ICLR 2026 🇧🇷 ELViS is a fast, lightweight, and interpretable module for estimating image-to-image similarity that generalizes well to many image domains. Paper: arxiv.org/abs/2603.28603 Code: github.com/pavelsuma/ELViS Come see poster today @ P4-#3715

TESO: Online Tracking of Essential Matrix by Stochastic Optimization Jaroslav Moravec, Radim Šára, Akihiro Sugimoto tl;dr: in title, for keeping the calibration of the stereo pair. #CVPR2026 oral arxiv.org/abs/2604.19420

Geometric Context Transformer for Streaming 3D Reconstruction Lin-Zhuo Chen, Jian Gao, Yihang Chen, Ka Leong Cheng, Yipengjing Sun, Liangxiao Hu, Nan Xue, Xing Zhu, Yujun Shen, Yao Yao, Yinghao Xu tl;dr: 1st img: keep all tokens, next: keep camera token only arxiv.org/abs/2604.14141

The picture I see, when I tell Claude "Let's go A"

"First clarifying question: What exactly do you mean by "DINOv3"? As of my knowledge cutoff (Aug 2025)..." Claude, are you OK today?

It seems, that we have failed the communication about IMC26. Let's try again. The competition this year is here: kaggle.com/competitions... No prizes, but whole year leaderboard -- similar to KITTY and other academic competitions. 3D people, please retweet and share.

Lyra 2.0: Explorable Generative 3D Worlds Tianchang Shen et 14 al. tl;dr: generate traj-controlled video -> make 3DGS from them, store and condition on them next. 18 pages paper. arxiv.org/abs/2604.13036

Human-in-the-Loop Local Corrections of 3D Scene Layouts via Infilling Christopher Xie, Armen Avetisyan, Henry Howard-Jenkins, Yawar Siddiqui, Julian Straub, Richard Newcombe, Vasileios Balntas, Jakob Engel tl;dr: well, in title. SceneScript is cool arxiv.org/abs/2503.11806

SpatialEdit: Benchmarking Fine-Grained Image Spatial Editing Yicheng Xiao et 12 al. tl;dr: official confirmation that NanoBanana is the🐐, but far from perfect. Also proposed QWEN3-based opensource model for the task arxiv.org/abs/2604.04911

What is happening with bsky servers and GCP now?

Multimodal Language Models Cannot Spot Spatial Inconsistencies Om Khangaonkar, Hadi J. Rad, Hamed Pirsiavash tl;dr: in title. The task is hard enough for humans as well - I have to spend 5-20 sec examining every object to spot the inconsistency. arxiv.org/abs/2604.00799

I'm making a cozy and relaxing town builder called 🌳🏡Island Architect 🏡🌳! - Coastal European vibe - Sculpt terrain, including caves & sea arches - Build wonky, curved buildings Wishlist now on Steam: store.steampowered.com/app/3079100/...

Repeat after me: supervision with Colmap does NOT mean "no supervision". No, supervising with MapAnything does NOT mean "relieving from GT annotation". It is not "self-supervision" either. Screenshots: Reliev3r (2026), SGP (2021).

LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, Ming-Hsuan Yang, Forrester Cole, Trevor Darrell, Deqing Sun arxiv.org/abs/2603.03269 tl;dr: TTT is new BA? new memory (for previous frames) representation for it.

MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data Clémentine Grethen, Yuang Shi, Simone Gasparini, Géraldine Morin tl;dr: synthetic lunar dataset, can be used for lunar image matching test/train. arxiv.org/abs/2604.00682

3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image Ze-Xin Yin, Liu Liu, Xinjie Wang, Wei Sui, Zhizhong Su, Jian Yang, Jin Xie tl;dr: in title, cool dataset. #CVPR2026 arxiv.org/abs/2604.04406

InCaRPose: In-Cabin Relative Camera Pose Estimation Model and Dataset Felix Stillger, Lukas Hahn, Frederik Hasecke, Tobias Meisen tl;dr: I don't know why you would need such thing, but interesting study for small displacement + wide lens anyway. arxiv.org/abs/2604.03814

Here's a quick thread about how the procedural material is set up to enhance the details of the editable terrain in Island Architect, without baking any textures. Still details I want to add, but I'm happy with the foundation! Here's the final result and the diffuse texture: #gamedev #indiegame

Anthropic: "Mythos is too dangerous". Claude Code, shortly after: "API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"}". Has machine rebellion already started?

Is it a csv, when the file has single column and no separator?