To have accuracy as "circle size" now in my top of misleading plot examples. Not to mention, that star doesn't obey even this strange idea
Dmytro Mishkin
@ducha-aiki.bsky.social
Marrying classical CV and Deep Learning. I do things, which work, rather than being novel, but not working. http://dmytro.ai
Agents finally brought the promise of all data labeling companies failed to deliver - you can label a little of data, new model/alg is trained on them and helps you more and more. The labeling is becoming adaptive as well. I love this
Having fun with Blender ... Distilling context transformers into recurrent transformers: arxiv.org/abs/2606.21562 @weinzaepfelp.bsky.social @mbsariyildiz.bsky.social
The best and the worst work for the agents is speed optimizations. Best - because it is well defined and easy to climb leaderboard. Worst - because you cannot use agent swarm, as the timings would be unreliable.
In addition to the research findings, this is the type of research I really enjoyed doing 😊 "On Hyperparameter Tuning on the Test Set" arxiv.org/abs/2610.05902
On Hyperparameter Tuning on the Test Set
"Don't tune hyperparameters on the test set" is often stated in machine learning textbooks. Violating it is considered a cardinal sin that produces misleadingly optimistic results, corrupts benchmark ...
arxiv.org
arXiv has updated our policy on rate limiting for all submitters. This update was made to fairly distribute moderator time & support the arXiv community of staff, volunteers, readers & authors. Please read our announcement to learn more: https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/
One of the aspects of industry chatbots that scares me the most. I have refused to work on persuasion enhancement since early in my PhD.
Bhaskar Mitra | ভাস্কর মিত্র@bmitra.bsky.social · 2w agoThe biggest threat that AI poses to our information ecosystem isn't so-called "hallucinations"; it's intentional large-scale persuasion / manipulation. bsky.app/profile/bmit...
🏆 The 3DV Citations Grand Prix! P2, with 3337 citations. "Matterport3D: Learning from RGB-D Data in Indoor Environments". arxiv.org/abs/1709.06158 Shout out to Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang!
Is it still true: is there trigonometry in US high schools
Visualization of the day: Slices through visual spacetime Interactive webpage: csprofkgd.github.io/spacetime-vo...
TIL - unlike Python - bash reads scripts incrementally. ``` echo "Starting..." sleep 30 # Change this before the sleep finishes, and the new message will be printed. echo "Finished!" ```
Example questions: (1) What papers look like after AI (2) What is would look like for a positive society-wide benefit from AI (3) What job training looks like if tasks become easier Trying to work through my own vision and it's nice to assemble others
A lesson from decades of work on automation is that autonomous systems require continual monitoring and human oversight. System designers never(*) succeed in anticipating all of the ways that a system can fail.
Amazing little feature on bsky -- automatic indicator of number tweets in the thread.
New, not so slop AI pop science video on YouTube. Kind of reasonable topic and not terrible delivery, didn’t even saw to “ai content” stamp. Next video topic helped me though: “Why we are eating tilapia and not carp?” WTF you mean not eating carp? Will stick to Veritasium
It is always tempting to add "consistency loss term" in 3D in one way or another. Do not do it. If it helps, then your data is too small, or you are predicting disconnected outputs. Just clf/regress vs ground truth, and do it long enough.
our latest work on instance-level recognition by learning from generated images is out! with @yankunwu.bsky.social, Zakaria Laskar, @gkordo.bsky.social and @gtolias.bsky.social
TMLR Published Papers@tmlr-pub.bsky.social · last mo.Instance-Level Generation for Representation Learning Yankun Wu, Zakaria Laskar, Giorgos Kordopatis-Zilos, Noa Garcia, Giorgos Tolias Action editor: Brian Kulis https://openreview.net/forum?id=T3JgJXH3ZK #datasets #recognition #instance
🔥 Composed image retrieval, as it should be. 📸 i-CIR (NeurIPS 2025): a photo says which particular object, a text says how it should look. 🔍︎ Explore every composed query in our new dataset browser: vrg.fel.cvut.cz/icir/ More to come... #ComputerVision #ImageRetrieval #AI #MultimodalSearch
i-CIR: Instance-Level Composed Image Retrieval (NeurIPS 2025)
A photo says which object, a text says how it should look. 202 instances, 1,883 composed queries, 752K curated hard negatives as hard as 40M random distractors, and an interactive dataset browser.
vrg.fel.cvut.cz
Sadly, I don’t monitor arXiv anymore. I don’t even read Scholar Inbox suggested papers, although I am still going to read VidMap carefully. If you subscribed to me for the arXiv feed only, probably unsubscribe.
In most practical cases this quickly happens automatically due to the quota system (all submissions count, including desk rejections) medium.com/@TmlrOrg/ann...
Annual Author Submission Quotas for TMLR
TMLR submissions have grown substantially in recent times (Figure 1), tripling in the past one year. Among these, the number of…
medium.com
Nihar Shah did a heroic experiment for TMLR: he spent 20-25 hours over two weeks interviewing authors of seemingly low-quality submissions about their own papers. He confirmed what we all suspected: people submitting these papers have *no idea* what is going on in them.
Transactions on Machine Learning Research@tmlrorg.bsky.social · 4w agoTMLR has faced a deluge of submissions, necessitating stricter desk rejection policies due to limited reviewer capacity Co-EiC Nihar Shah reached out to authors of 10 papers slated for desk reject. Could they answer questions about their *own* submission? medium.com/@TmlrOrg/ask...
gotta keep tapping the sign peoples-things.ghost.io/why-do-peopl...
Against Narcissistic-Sociopathic Technology Studies, or Why Do People *USE* Technologies?
Why and how do people USE technologies? This question should be at the center of any thinking about human life with technology. The ways in which people use technologies, after all, determines the so...
peoples-things.ghost.io
(1/n) Can older lidar SSC models catch up with recent SOTA just by getting better inputs? At #ICIP2026 today and in our new journal extension, we add two cheap priors to existing networks, with no redesign. 📍 Poster #2752, Park Hall, 14:00–16:00 🇫🇮 🧵👇
Here is a thread of all the good reasons to use outlook! 🧵👇
#ICIP2026 First impressions: the venue is amazing! If you’re attending and wanna talk all things 3D scene understanding - hit me up on the conference app!
Yeah the booklet would have been a lot better. Conference apps can be good, but Cvent is not. It's slow, the UI is terrible and it crashes on my phone when listing sessions. And the ECCV website is not better. 🤷🏻♂️
ECCV Poster #305 at 3pm (now): A scalar per patch from pre-trained ViTs enables fast moving navigation in the real world @steevenj7.bsky.social, L. Antsfeld, myself (I will present) - Visual enc distilled from heterog. teachers - Attention proj - Pretrain Policy w Lidar input => remove Lidar
A robot moves hook from A to B and we measure battery drain (assuming perfect operation). Hook catches 50kg bag of rice. Stupid robot continues to pull hook and bag, while clever one untangles and proceeds w/o bag. Is there metric/formula for taking stupidity into account?