Dima Damen

The hidden flaws in your favorite foundation model: We've uncovered how subtle image metadata (JPEG params, camera type, etc.) systematically biases visual representations and consequently affect the object recognition ability. To be presented at #ICCV2025 as a highlight paper.

No we shouldn't... but also every metric can be optimised through shortcuts and cheating. That doesn't make an evaluation metric itself useless.
The h-index is only meaningful to compare one's career over time, not across people. When calculated (without cheating), it shows a useful trajectory.

not even for tasks such as stirring your food or washing dishes? :-)

This is one of the coolest ideas using EPIC-KITCHENS in a long while...
We've all been waiting to be replaced by robots! At least this is now done in the generative space...
Great work by Marion Lepert, Jiaying Fang and @leto--jean.bsky.social from Stanford IPRL.. congrats!
arxiv.org/abs/2508.09976

Vision is gen-NI !

('The shape of things unseen', Adam Zeman)

Happy Birthday Kosta… thanks for sharing the lovely pictures!
Wishing you and your adorable family the next 50 years of Happiness, Heath and Success 🥳🥳

We have released the winners list and winning reports for the 2025 EPIC-KITCHENS and HD-EPIC VQA Challenges, awarded at the 2nd #EgoVis workshop @cvprconference.bsky.social #CVPR2025
Check these reports:
epic-kitchens.github.io/2025#results
hd-epic.github.io/index#vqa-be...

After a round of mini-golf!
@prajwalgatti.bsky.social @zhifan-zhu.bsky.social @sinhasaptarshi.bsky.social Siddhant Bansal, WeiHong Li, Ahmad Darkhalil, Rhodri Guerrier, Fahd Abdelazim, Sam Pollard, Masatoshi Tateno

2/2

Group socials are key to bringing people together, reminding us of the person behind the work and wishing those leaving all the best for their future career...
We #MaVi @bristoluni.bsky.social met to wish postdoc Kranti Parida, leaving us this week all the best,
At Bristol Balloon Festival
1/2

BBC Inside Science BBC Radio 4 has been exploring the most powerful computer the UK has ever seen 🖥️🤯

Hear how our Isambard-AI #supercomputer is being used to carry out groundbreaking new research: www.bbc.co.uk/sounds/play/...

Finally the next @iclr-conf.bsky.social location is revealed...
iclr.cc
#ICLR2026 will be in Rio de Janeiro from 23 to 27 April!

If you have not seen this yet, you are missing a lot!
Genie 3 by Google DeepMind was unveiled today &delivers in abundance.
Of course my fav example is ego x world model.
It is video gen x modeling "out of the frame".
Many congrats @jparkerholder.bsky.social & team
deepmind.google/discover/blo...

Pascal VOC 2007 Dataset - Machine Learning Datasets Load the Pascal VOC 2007 dataset in Python fast. 20 object classes, 9,963 images, with 24,640 labeled samples. Stream Pascal VOC 2007 while training ML models.

Actually the one you need is the 2007 one where the pixel masking task was defined but the link appears broken to the report..?
Maybe you can find it in other ways,
datasets.activeloop.ai/docs/ml/data...

You should read the Pascal POC challenge definition
link.springer.com/article/10.1...

Dark blue background to the left of image with Cardiff University logo in top left corner, red Babylab dragon and ident to the centre. Cream box overlays the blue background. Centred dark blue text reads "are you a medical or allied health professional based in the UK? Do you work with children with Down syndrome under 5? We want to hear your thoughts on using wearable head-mounted cameras in your practice". A QR code to the bottom left with dark blue text that reads "Find out more about our new remote study using the QR code or link in the caption" Right side of image shows a young child with Down syndrome in a light blue t-shirt with the Babylab Dragon and text reading “Little Scientist”. She is wearing a soft, light blue foam helmet with a small camera attached to the front, just above eye level.

Cardiff Babylab is excited to launch a new, remote study for medical and allied health professionals in the UK!

We are inviting professionals to give their thoughts on integrating wearable head cameras into their practice.

Find out more: www.cardiff-babylab.com/tinyexplorer...

Many thanks @pascalmettes.bsky.social @ellisamsterdam.bsky.social for visiting us @bristoluni.bsky.social to examine (now Dr) Adriano Fragomeni (supervised by myself and Michael Wray) and give a great talk on hyperbolic deep learning. Enjoyed your visit

Extended EPIC-SOUND paper was accepted at TPAMI
arxiv.org/abs/2302.006...
This follows ICASSP 2023 oral, extended for detection and further analysis
epic-kitchens.github.io/epic-sounds/
work by @jaesunghuh.bsky.social Jacob Chalk @ekazakos.bsky.social
@oxford-vgg.bsky.social @bristoluni.bsky.social

wouldn't you be distilling some intermediate features rather than just the label?
No clue, I've never worked on distillation myself

Perception Test [3rd edition]
perception-test-challenge.github.io
github.com/google-deepm...
led by Viorica Patraucean, Joe Heyward, @nikparth1.bsky.social @tylerzhu.bsky.social
João Carreira, AZ and myself

Up to 50K in prizes sponsored by Google DeepMind

More on PT:
youtu.be/8BiajMOBWdk?...
4/4

- two guest tracks:
- KiVA image understanding challenge
kiva-challenge.github.io
‪@euniceyiu.bsky.social @shiryginosar.bsky.social

- Physics-IQ video generation challenge
physics-iq.github.io/workshop/phy...

3/4

Our new novel tracks unify diverse tasks under common interfaces to move beyond single-task models:
- joint object/point tracking
- joint action/sound localisation
- unified multiple-choice videoQA
Also:
- novel VLM interpretability track -- can you show where models fail?
2/4

Join us for 3rd Perception Test Workshop &Challenge
@iccv.bsky.social #iccv2025
*NEW* this year:
- 3 unified tracks
- novel interpretability track
- guest tracks: KiVA and Physics-IQ
- 4 world-class speakers (see pic)
Up to 50K in prizes sponsored by Google DeepMind
🧵 for details [1/4]

Scaling 4D Representations

Scaling 4D Representations

Self-supervised learning from video does scale! In our latest work, we scaled masked auto-encoding models to 22B params, boosting performance on pose estimation, tracking & more.

Paper: arxiv.org/abs/2412.15212
Code & models: github.com/google-deepmind/representations4d

Another poster presented yesterday at ICVSS is our collaboration with Keio University in Japan. Masashi Hatano spent 1-year visiting @bristoluni.bsky.social and was presenting our ArXiv paper: The Invisible EgoHand at #ICVSS2025. Great work worth your attention
arxiv.org/abs/2504.08654
3/3

Also proud of the HD-EPIC team (Omar, Fahd, Kaiting) who are attending @ICVSS on behalf of @bristoluni.bsky.social and @unileiden.bsky.social. They had a busy poster session #ICVSS2025 yesterday detailing our @cvprconference.bsky.social work and public dataset
hd-epic.github.io
2/3

Giving a tutorial at #ICVSS2025 is special - you could inspire the brightest upcoming researchers in the field, eager to learn and discuss.
I talked about a new passion of mine - Video Understanding *out of the frame*.
If interested slides are at: tinyurl.com/Dima-ICVSS2025
🧵1/3

How to make 3D fundamental to all aspects of computer vision?
A Vedaldi @oxford-vgg.bsky.social with a brilliant tutorial this morning #icvss2025

No no... really my 40th birthday :-D [which was 4 years ago]. I invited my friends to "Go Ape" which is what these are called in the UK :-)
and we really had fun!

Love them too, including the zip lines at the end. I went there for my 40th 🤭

In a new paper led by Gianluca Monaci, with @weinzaepfelp.bsky.social and myself, we explore the relationship between rel pose estimation and image goal navigation and study different architectures: late fusion, channel cat (w/ or w/o space2depth) and cross-attention.

arxiv.org/abs/2507.01667

🧵1/5