Yannis Kalantidis

Annie and Lakeesha struggle in school. AI teacher assistants treated them very differently. A Common Sense Media study found that prominent teacher assistants that use AI generated recommendations that appeared to be rooted in racial stereotypes based on students’ names. About a third of tea...

"Asked to generate intervention plans for struggling students, AI teacher assistants recommended more-punitive measures for hypothetical students with Black-coded names and more supportive approaches for students the platforms perceived as white" www.chalkbeat.org/2025/08/06/a...

most people want a quick and simple answer to why AI systems encode/exacerbate societal and historical bias/injustice and due to the reductive but common thinking of "bias in, bias out," the obvious culprit often is training data but this is not entirely true

1/

The biggest update in 3D reconstruction world - VGGT (new weights) has been released for commercial usage as well.
Kudos to @jianyuanwang.bsky.social to make this happen!
github.com/facebookrese...

AI and Fraternity, Abeba Birhane, AI Accountability Lab I envision a future where human dignity, justice, peace, kindness, care, respect, accountability, and rights and freedoms serve as the north stars that guide AI development and use. Realising these ideals can’t happen without intentional tireless work, dialogues, and confrontations of ugly realities – even if they are uncomfortable to deal with. This starts with deciphering hype from reality. Pervasive narratives portray AI as a magical, fully autonomous entity approaching a God-like omnipotence and omniscience. In reality, audits of AI systems reveal a consistent failure to deliver on grandiose promises and suffer from all kinds of shortcomings, issues often swept under the rug. AI in general, and GenAI in particular, encodes and exacerbates historical stereotypes, entrenches harmful societal norms, and amplifies injustice. A robust body of evidence demonstrates that — from hiring, welfare allocation, medical care allocation to anything in between — deployment of AI is widening inequity, disproportionately impacting people at the margins of society and concentrating power and influence in the hands of few. Major actors—including Google, Microsoft, Amazon, Meta, and OpenAI—have willingly aligned with authoritarian regimes and proactively abandoned their pledges to fact-check, prevent misinformation, respect diversity and equity, refrain from using AI for weapons development, while retaliating against critique. The aforementioned vision can’t and won’t happen without confrontation of these uncomfortable facts. This is precisely why we need active resistance and refusal of unreliable and harmful AI systems; clearly laid out regulation and enforcement; and shepherding of the AI industry towards transparency and accountability of responsible bodies. "Machine agency" must be in service of human agency and empowerment, a coexistence that isn't a continuation of modern tech corporations’ inequality-widening,

so I am one of the 12 people (including the “god-fathers of AI”) that will be at the Vatican this September for a two full-day working group on the Future of AI

here is my Vatican approved short provocation on 'AI and Fraternity' for the working group

Deadlines for ICLR 2026 have been communicated on the official site:
Abstract: Sep 19 '25 (Anywhere on Earth)
Paper: Sep 24 '25 (Anywhere on Earth)
iclr.cc/Conferences/...

ICLR 2026 will happen in Brazil! ❤️

(Don't fall for the predatory conferences with the same name)

New research from MIT found that those who used ChatGPT can’t remember any of the content of their essays.

Key takeaway: the product doesn’t suffer, but the process does. And when it comes to essays, the process *is* how they learn.

arxiv.org/pdf/2506.088...

When were #CVPR2025 papers available on arXiv? 👇

#CVPR2025 attendance per country, comparing 2025 and 2024, sorted by diff. Numbers were extracted from these two slides presented by the respective program chairs.

(Totals on the slide seem to include virtual registrations, but not the per country numbers)

Diff between western and eastern Europe?

I have created a starter pack with researchers from Naver Labs Europe @naverlabseurope.bsky.social: we are in Grenoble, France, and we do research in AI for robotics, computer vision, NLP, machine learning, HRI.
go.bsky.app/JdTFu4Q

Our work on "Reasoning in visual navigation..." presented as a "Highlight" by Boris Chidlovskii and Francesco Giuliari at #cvpr2025!

Interactive site, play around with dynamical models:
europe.naverlabs.com/research/pub...

Thanks @weinzaepfelp.bsky.social for the photo.

@steevenj7.bsky.social

During today's #CVPR2025 workshops, I will present:
- What matters in ImageNav: architecture, pre-training, sim settings, pose (poster & highlight at the Embodied AI workshop)
- CondiMen: Conditional Multi-person Human Mesh Recovery (Poster at the Rhobin workshop and at the 3D Humans workshop)

Paper: arxiv.org/abs/2503.19777
Code: github.com/vladan-stojn...
Demo: huggingface.co/spaces/stojn...

Work with @skamalas.bsky.social, Jiri Matas, and @gtolias.bsky.social.

Wanna the outstanding performance of MASt3R while using a ViT-B or ViT-S encoder instead of its ViT-L one? Don't miss how we build DUNE, a single encoder for diverse 2D & 3D tasks, at this afternoon #CVPR2025 poster session (poster #376).

paper: arxiv.org/abs/2503.14405
code: github.com/naver/dune

Are you at @cvprconference.bsky.social? Come by our poster!
📅 Sat 14/6, 10:30-12:30
📍 Poster #395, ExHall D

💜 Huge thanks to our amazing organizers for making #WiCV
@cvprconference.bsky.social 2025 possible! 🙌

After 6+ months of near-daily work by 9 dedicated organizers, it’s exciting to see everything come together.
So proud of this amazing team! 💪✨

wait a minute... is this the Acro... :P

You woke up early in the morning jet-lagged and having a hard time deciding for a workshop today @cvprconference.bsky.social ?

Here's a reliable choice for you: our workshop on 🛟 Uncertainty Quantification for Computer Vision!

🗓️ Day: Wed, Jun 11
📍Room: 102 B
#CVPR2025 #UNCV2025

Have you heard about HD-EPIC?
Attending #CVPR2025
Multiple opportunities to know about the most highly-detailed video dataset with a digital twin, long-term object tracks, VQA,…
hd-epic.github.io

1. Find any of the 10 authors attending @cvprconference.bsky.social
– identified by this badge.

🧵

Check out our new paper in #CVPR2025! We test the limits of multi-teacher distillation to get one of the most versatile encoders yet. Mixing DINOv2 semantics 🦕, MASt3R multi-view reconstruction 📸📸...📸 and Human Mesh Recovery 🕺💃...🏌️‍♂️
Check our project page: europe.naverlabs.com/research/pub...

10/ 💬 Both papers tackle generalization head-on —
One by building a universal encoder from heterogeneous knowledge, the other by refining language-driven predictions without training.

Don't forget to check them out at #CVPR2025 in Nashville! 🎸

GitHub - vladan-stojnic/LPOSS: Code for LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation (CVPR2025) Code for LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation (CVPR2025) - vladan-stojnic/LPOSS

9/ 📈 Results:
→ Substantial mIoU boost on 8 datasets
→ LPOSS+ especially shines for boundary precision
→ Outperforms many VLM-based methods without training

Work w/ Vladan @stojnicv.xyz and Giorgos @gtolias.bsky.social from CTU in Prague
Code: github.com/vladan-stojn...
Demo: hf.co/spaces/stojn...

8/ 🔁 LPOSS stages:
🔸 LPOSS: Refines patch-level predictions
🔹 LPOSS+: Refines on pixel-level using upsampled LPOSS output
✅ Uses DINO features + spatial cues for affinity
✅ Totally training-free!

7/ 🚫 VLMs are great at training-free segmentation, but:

* Predictions are patch-based (low-res)
* Predictions are performed independently at each patch (noisy)

LPOSS improves this with test-time label propagation, without fine-tuning.

6/ 📄 Paper 2:
"LPOSS: Label Propagation Over Patches and Pixels for Open-Vocabulary Semantic Segmentation"

Can graph-based label propagation refine weak, patch-level predictions from VLMs like CLIP? We say yes — introducing LPOSS and LPOSS+.

DUNE: Distilling a Universal Encoder from heterogenous 2D and 3D teachers CVPR 2025 publication

5/ Project page: europe.naverlabs.com/dune

Work with Bulent Sariyildiz @mbsariyildiz.bsky.social Diane Larlus @dlarlus.bsky.social Philippe Weinzaepfel @weinzaepfelp.bsky.social Pau de Jorge @pdejorge.bsky.social and Thomas Lucas at NAVER LABS Europe @naverlabseurope.bsky.social

GitHub - naver/dune: Code repository for "DUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D Teachers" Code repository for "DUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D Teachers" - naver/dune

4/ 💡 Results:
✅ Strong performance on diverse tasks
✅ #1 on Niantic's Map-free Relocalization leaderboard, inproving over MASt3R with a smaller encoder (ViT-Base vs. ViT-Large)

Code and models: github.com/naver/dune

3/ 🔍 DUNE is trained from heterogeneous teachers using heterogeneous data:
🟧 DINOv2 (2D foundation model - Oquab et al @ Meta)
🟥 MAST3R (3D foundation model - @vincentleroy.bsky.social et al @ NAVER LABS Europe - )
🟨 Multi-HMR (3D human perception - @fbaradel.bsky.social et al @ NAVER LABS Europe)

2/ 🧠 Why DUNE?
→ 2D & 3D vision models (e.g., segmentation, mono- or binocular depth, human mesh, 3D reconstruction) are siloed.
→ Can we distill their knowledge into a single universal encoder?
Yes — using multi-teacher distillation with transformer projectors, teacher dropping, and more.

1/ 📄 Paper 1:
"DUNE: Distilling a UNiversal Encoder from Heterogeneous 2D and 3D Teachers"
We propose DUNE: a ViT-based encoder distilled from multiple specialized 2D & 3D foundation models to unify visual tasks across 2D, 3D and human understanding.

🧵 Two new papers at #CVPR2025 on generalization in visual representations — covering both universal encoders for 2D and 3D tasks, and open-vocabulary semantic segmentation

Let's dive in! 👇