🚀New task: Instance-level Image+Text→Image Retrieval

🔎Given a query image + an edit (“during night”), retrieve the same specific instance after the change — not just any similar object.

🛢New dataset on HF: i-CIR huggingface.co/datasets/bil...

🔥Download, run, and share results!

Instance-Level Composed Image Retrieval The progress of composed image retrieval (CIR), a popular research direction in image retrieval, where a combined visual and textual query is used, is held back by the absence of high-quality training...

⬇️ Grab i-CIR, run your method, tell us how it handles instance-level composed image retrieval.

📄 arxiv.org/abs/2510.25387
🧪 github.com/billpsomas/i...

George Retsinas, @nikos-efth.bsky.social, Panagiotis Filntisis, Yannis Avrithis, Petros Maragos, Ondrej Chum, @gtolias.bsky.social.

Instance-Level Composed Image Retrieval
@billpsomas.bsky.social George Retsinas @nikos-efth.bsky.social Panagiotis Filntisis,Yannis Avrithis, Petros Maragos, Ondrej Chum, @gtolias.bsky.social

tl;dr: condition-based retrieval (+dataset) - old photo/sunset/night/aerial/model arxiv.org/abs/2510.25387

Pattern Recognition and Computer Vision Colloquium - past speakers

The Colloquium in Pattern Recognition and Computer Vision of the Visual Recognition Group at CTU in Prague has a long tradition dating back to 1998. The list of all speakers is available docs.google.com/spreadsheets.... Enjoy! The 50th edition is coming soon cmp.felk.cvut.cz/colloquium/

🚨 Deadline Extension

Instance-Level Recognition and Generation (ILR+G) Workshop at ICCV2025 @iccv.bsky.social

📅 new deadline: June 26, 2025 (23:59 AoE)
📄 paper submission: cmt3.research.microsoft.com/ILRnG2025
🌐 ILR+G website: ilr-workshop.github.io/ICCVW2025/

#ICCV2025 #ComputerVision #AI

Exciting News!
Shoutout to @hf.co 🤗 for hosting ILIAS in its full scale!

You can now easily download it via:
huggingface.co/datasets/vrg...

This marks a milestone in i2i and t2i retrieval, paving the way for significant advancements in the field.

Huge thanks to @hf.co team for their support!

The WACV'25 work of @nikos-efth.bsky.social is on WACV daily - computer vision news.

ILIAS is a large-scale test dataset for evaluation on Instance-Level Image retrieval At Scale. It is designed to support future research in image-to-image and text-to-image retrieval for particular objects and serves as a benchmark for evaluating foundation models and retrieval techniques.

ILIAS: Instance-Level Image retrieval At Scale

@gkordo.bsky.social, Vladan Stojnić @annetka.bsky.social Pavel Šuma, Nikolaos-Antonios Ypsilantis @nikos-efth.bsky.social Zakaria Laskar,Jiří Matas, Ondřej Chum, @gtolias.bsky.social

tl;dr: SigLIP rules. Lots of ablations
arxiv.org/abs/2502.11748
1/

For PhD and MSc students interested in a research visit to Prague/VRG in 2025: we're open to hosting short-term collaborations or internships on a range of computer vision topics. If this sounds exciting, reach out by e-mail! We'd love to discuss potential projects. Some examples 🧵
#Internship #CV

The most important latex command

Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization Single-source domain generalization attempts to learn a model on a source domain and deploy it to unseen target domains. Limiting access only to source domain data imposes two key challenges - how to ...

8/ Dive deeper here:
📖 Paper: www.arxiv.org/abs/2409.19774
💻 Code: github.com/NikosEfth/cr...

7/ Thanks to my supervisors and co-authors @gtolias.bsky.social and Ondrej Chum for making this possible!

[WACV 2025 Oral] Crafting Distribution Shifts for Validation and Training in Single-Source DG YouTube video by Nikos Efthymiadis

6/ Check out our WACV resources:
🎥 Video presentation: youtu.be/ZPw3CUioMMg
🖼️ Poster: cmp.felk.cvut.cz/~efthynik/po...
🌐 Project homepage: cmp.felk.cvut.cz/~efthynik/cr...

5/ Plus, we introduce a state-of-the-art single-source-domain-generalization method, setting new records on four popular generalization benchmarks.

4/ Our work offers a solution:
🌟 Augmenting the validation set! We achieve correlations of 0.74 (PACS) and 0.83 (Mini-DomainNet)—enabling confident model selection without access to the target domain!

3/ But what if your validation-to-test correlation looked random (left scatter plots)? 😬

2/ Here’s the setup:
Imagine building a model to classify photos 📸, but your goal is to generalize to sketches ✏️ or paintings 🖌️.

You only have labeled photos. The target domains are inaccessible due to cost or legal rights. You validate models on the source domain and hope for the best.

🧵 1/ Excited to share that our WACV 2025 paper, "Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization," was also selected for an Oral Presentation! 🎤

Composed Image Retrieval for Training-Free Domain Conversion This work addresses composed image retrieval in the context of domain conversion, where the content of a query image is retrieved in the domain specified by the query text. We show that a strong visio...

6/ Want to dive deeper?
📖 Paper: arxiv.org/abs/2412.03297
💻 Code: github.com/NikosEfth/fr...

#WACV2025 #ComputerVision #AI

5/ A huge thanks to my amazing co-authors:
@billpsomas.bsky.social, Zakaria Laskar, Konstantinos Karantzalos, Yannis Avrithis, Ondrej Chum, and @gtolias.bsky.social .

[WACV 2025 Oral] Composed Image Retrieval for Free Domain Conversion YouTube video by Nikos Efthymiadis

4/ check out our resources:
🎥 Video presentation: youtu.be/l57-rGT6zEs
🖼️ Poster: cmp.felk.cvut.cz/~efthynik/po...
🌐 Project homepage: cmp.felk.cvut.cz/~efthynik/fr...

3/
🖼️ We expand ImageNet-R to include more domains as image queries and introduce three new benchmarks, creating a more stable research base.

💡 We propose FreeDom: a novel, training-free approach, using the concept of memory-based textual inversion. 🚀

2/ The challenge: Domain conversion 🎨 is an underexplored task. Most efforts have focused on photo-to-rest conversions, like ImageNet-R. But our work expands the scope in two major ways:

🧵 1/ Big news! Our WACV 2025 paper, "Composed Image Retrieval for Training-Free Domain Conversion", has been selected for an Oral Presentation! 🎤✨

A late introduction: I am an Associate Professor at CTU in Prague and a computer vision researcher at the Visual Recognition Group (VRG - vrg.fel.cvut.cz). VRG consists of about 40 members and I am leading a small team within VRG. I am made in Greece and exported to France + Czech Republic 🧵⬇️

Excited to present UDON at NeurIPS '24 tomorrow (Thursday 12/12)! If you are interested in a scalable training method for multi-domain image embeddings, come to poster #1410 in the East Exhibit Hall A-C of the Vancouver Convention Center from 11 am to 2 pm (PST) to discuss!

Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization Single-source domain generalization attempts to learn a model on a source domain and deploy it to unseen target domains. Limiting access only to source domain data imposes two key challenges - how to ...

4/4
Also, we introduce a single-source-domain-generalization method, which is the new state-of-the-art on 4 benchmarks.

Our paper on ArXiv: www.arxiv.org/abs/2409.19774
Our code on GitHib: github.com/NikosEfth/cr...

Shout-out to my supervisors and co-authors @gtolias.bsky.social and Ondrej Chum

3/n
It offers a practical alternative to this problem! With domain-agnostic augmentations, we achieve correlations of 0.74 (PACS) and 0.83 (Mini-DomainNet)—letting you make confident model choices even without target domain access.

2/n
But what if the validation-to-test correlation looked like the scatterplots on the left? 😬

They are random at best, aren't they?

We are happy to present our second WACV 2025 paper named: "Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization".