VGGT-Ω @jianyuanwang.bsky.social et 9 al. tl;dr: 1) simplicity to scale compute(no pcl head, no matching head, but losses kept, no rays for intrinsics, less DPT) 2) Complex video SfM/filtering procedure to scale data. 3) works great on IMC-2025(not in paper, I tested) arxiv.or…
Who's Who of AI
Dmytro Mishkin
Why they matter
Directory member with public evidence across AI research.
- AI signals
- 3 past 30d
- Sources
- 1 distinct domains
- Discussões
- 3 past 30d
- Latest signal
- 5d ago
Marrying classical CV and Deep Learning. I do things, which work, rather than being novel, but not working.
http://dmytro.ai
What they're sharing
VGGT-$Ω$ arxiv.org
A Fix for Softmax and Cross-Entropy arxiv.org
Glob3R: Global Structure-from-Motion with 3D Foundation Models arxiv.org
Doomed to Re-Annotate, Forever: The ImageNet Story arxiv.org
FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation arxiv.org
Robust Local Optimization Done Right arxiv.org
arxiv.org
arxiv.org
Regret Optimal Control for Uncertain Stochastic Systems arxiv.org
DF26: We Cannot Tell Fake From Real Anymore arxiv.org
DF26/DF26 · Datasets at Hugging Face huggingface.co
thecv… openreview.net
Articles & links
Here you go arxiv.org/abs/2610.31642
AI Weekly's analysis
→
- The paper names two numerical bugs in standard softmax, log-softmax and cross-entropy: Softmax Collapse and Softmax Cancellation.
- The authors say their fix eliminates the slingshot phenomenon and speeds up grokking when weight decay is applied.
- In knowledge distillation, the corrected cross-entropy stops student networks from drifting away from the teacher, according to the abstract.
Read full analysis →
Glob3R: Global Structure-from-Motion with 3D Foundation Models Junyuan Deng et 9 al. tl;dr: finetune Pi3 matching head in RoMav2 style, use SALAD to sort seqs, do Glomap-like global SfM arxiv.org/abs/2607.09225
AI Weekly's analysis
→
- Glob3R augments a frozen Pi3X 3D foundation backbone with a lightweight dense matching head that predicts image warps between frames.
- Those warps feed sparse multi-view feature tracks into global motion averaging and bundle adjustment across large unordered image collections.
- The authors report gains over feed-forward foundation baselines on indoor, outdoor, driving, and unordered SfM benchmarks, plus better neural rendering.
Read full analysis →
Doomed to Re-Annotate, Forever: The ImageNet Story Illia Volkov, Nikita Kisel, Tetiana Mishkina, @klara-cz.bsky.social Jiri Matas tl;dr: ImageNet val annotation with all bboxes, careful study of concepts, attributes (is drawing of a cat, a cat?) Year of efforts. 1/ arxiv.org/a…
FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation P.Lindenberger @pesarlin.bsky.social O.Miksik @marcpollefeys.bsky.social tl;dr: transformer for dinov3 token matching, query with anything (grid, keypoints) + image overlap clf. Trained with negatives! arxiv.…
Robust Local Optimization Done Right James Pritts, Kevin Köser tl;dr: a Welsch robust loss with half-quadratic optimization, or how to use outliers to help. Sota in inl_th robustness vs SIMFIT++. also - RANSAC in Julia arxiv.org/abs/2610.05743 github.com/VisualGeomet...
Some work on similar, but for algorithmic case, when you kind of want to do it. arxiv.org/pdf/2503.13433 Earlier work from ours, arxiv.org/pdf/2003.01587 The issue is not tuning on test alone - it is tuning for YOUR method, crappy baseline of other - every line crossing is opp…
Some work on similar, but for algorithmic case, when you kind of want to do it. arxiv.org/pdf/2503.13433 Earlier work from ours, arxiv.org/pdf/2003.01587 The issue is not tuning on test alone - it is tuning for YOUR method, crappy baseline of other - every line crossing is opp…
That’s a regret, ok. arxiv.org/html/2304.14...
DF26: We Cannot Tell Fake From Real Anymore Severyn Shykula,Andrii Yermakov,Ivan Samarskyi, @ducha-aiki.bsky.social Jan Cech,Anastasiia Mishchuk tl;dr: "public speaking" dataset, recent vidgen. Humans cannot reliably detect fakes. No sound arxiv.org/abs/2609.07369 huggingface.…
AI Weekly's analysis
→
- On the DF26 benchmark, humans identified synthetic public-speaking video at 52.6 percent, versus 74.5 and 69.8 percent on two earlier deepfake datasets.
- The best detector tested, GenD-PE, reached only 69.7 AUROC, while two temporal detectors dropped to 48.2 and 61.6 AUROC on DF26.
- DF26 covers 271 real and 2,420 synthetic clips from seven generators including Veo 3.1, Kling 3.0, Wan 2.6, and HunyuanVideo 1.5.
Read full analysis →
DF26: We Cannot Tell Fake From Real Anymore Severyn Shykula,Andrii Yermakov,Ivan Samarskyi, @ducha-aiki.bsky.social Jan Cech,Anastasiia Mishchuk tl;dr: "public speaking" dataset, recent vidgen. Humans cannot reliably detect fakes. No sound arxiv.org/abs/2609.07369 huggingface.…
Structure from Motion in Deep Learning Era - save Wednesday September 9 morning! 3 speakers, 8 papers. sfm-dl.fbk.eu Papers are at OpenReview: openreview.net/group?id=thecv… #ECCV2026 @eccv.bsky.social
Their own posts
Recent commentary
Tomorrow is a better day...because USM3D workshop at @cvprconference.bsky.social is happening? 6 keynote speakers. 2 challenges which I am finally not ashamed of? Debatable. Is it our final competition at HuggingFace? The only way to figure out is to attend. Mile High 3B. #CVPR2026
Using generated AI “photos” in restaurant/ food delivery menus/recipe websites is lying to your customers/users. Using stock photos is lying to your customers. Just photo YOUR damn dish. You can even “enhance” it with AI, but at least ingredients, portions, texture stays real.
From the org/workshop PC, CMT is miles above OpenReview. Literally, I cannot do anything w/o asking ChatGPT how to do it. Never happened with CMT.
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
Just thought in EU I was safe from "tax separate" idiocy, but here go ChatGPT/Claude invoices with VAT separately. I guess I have to be happy that I don't need to add tip on top of that :)
Their network
In Dmytro Mishkin's orbit
Center = Dmytro Mishkin. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Dmytro Mishkin? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/ducha-aiki-bsky-social)