Let me introduce our new paper: Multimodal Large Language Models as Image Classifiers
β Multimodal LLMs are increasingly used for visual tasks, but evaluating their image classification ability has produced conflicting conclusions.
Link: arxiv.org/html/2603.06...
We validate EP across diverse pretrained backbones. It complements LoRA tuning and delivers improved object localization through its internal attention maps.
Simple idea. Strong gains. Broad applicability.
arxiv: arxiv.org/pdf/2506.10178
EP becomes especially effective when the backbone is pretrained for local representation learning, such as MAE. If your downstream task requires global prediction, EP bridges that gap. MAE-style models can, in fact, excel at global tasks when paired with the right probe.
Efficient Probing will be presented at ICLR 2026.
We introduce EP, an attentive probing method that consistently outperforms linear probing and prior attentive approaches. It's a simple, intuitive design that avoids over-parameterization compared to the black-box use of standard components.
New proceedings means low chance to be indexed by scopus/web-of-science from its first year, with the consequence of not getting recognized by some grant agencies, for example in Czech Republic. I recall the NeurIPS datasets track was not indexed from year 1.
The new CTU Rector begins their term in office with strong support for excellence. CTU has just launched a Starting Grant to attract outstanding earlyβcareer researchers who wish to join CTU and establish their own research group. Funding: up to β¬160k per year for 3 years. Deadline: 30 March 2026.
Clarifications for eligibility: 3 papers in total with each one being either a CORE A*/A conference or a journal with IF.
Start date is negotiable. Gross salary is 75 000 CZK. Plus the possibility of up to 20% extra in bonuses.
I have an opening for a two years post-doc position on instance-level (personalized) visual generation. Eligibility: (i) <=7 years from Ph.D. (ii) studies or 1 year outside of Czechia (ii) >=3 journal with IF or CORE A*/A conference papers. Deadline: 15 Feb.
Details: www.euraxess.cz/jobs/399390
πNew task: Instance-level Image+TextβImage Retrieval
πGiven a query image + an edit (βduring nightβ), retrieve the same specific instance after the change β not just any similar object.
π’New dataset on HF: i-CIR huggingface.co/datasets/bil...
π₯Download, run, and share results!
π£ i-CIR dataset (NeurIPS 25) is now on
@hf.co.
πEasier download + better discoverability + WebDataset shards for large-scale use (~750K images).
π€ Grab it here: huggingface.co/datasets/bil...
#computervision #retrieval #datasets #huggingface #NeurIPS
1/n REGLUE Your Latents! π
We introduce REGLUE: a unified framework that entangles VAE latents β Global β Local semantics for faster, higher-fidelity image generation.
Links (paper + code) at the endπ
This is a very serious initiative. While AGI risk debates get much attention, we should worry more about the immediate danger from AIβs role in automating war and surveillance.
maybe it's time for a larger cvpr in Paris?
It was a big pleasure to be in Nicolas's committee. Congratulations to Nicolas for the great work, and congratulations to the advisors too!
Prof. @tokehoye.bsky.social (Aarhus University) and I have an open PhD position (jointly advised) on biodiversity monitoring with camera trap networks. Deadline: 15-Jan-2026
Please help us share this post among students you know with an interest in Machine Learning and Biodiversity! π€πͺ²π±
This is a paper that will be presented next month at #NeurIPS2025. The dataset and code are already publicly available.
The studied setting allows to explore large image collections in flexible and creative ways: query with an image showing a particular object and add a text query to transform aspects like context, environment, lighting conditions, object state, and more.
There is a lot of work done recently on composed image retrieval, but we felt that none of the existing benchmarks reflect the real-world challenges and applications. So, we created a new test benchmark for instance-level composed image retrieval.
Looking for a PhD program? It all starts with great supervision. Choose wisely.
www.nature.com/articles/d41...
AnyUp is great. We are already using it flawlessly.
This is the real-world harm of computer vision: false accusations of gun possession over a bag of crisps. Deploying this in public is reckless.
www.bbc.com/news/article...
Honored to receive a Google award to support research on vision-language models for retrieval. Grateful for the opportunity to strengthen our collaboration with Google researchers, especially Ahmet Iscen.
All slides for the RANSAC in 2025 tutorial are online
#ICCV2025
danini.github.io/ransac-2025-...
Today at #ICCV2025 (afternoon poster session): see how sensitive some foundational models are to non-semantic cues like JPEG compression and camera model. Such cues can heavily distort their semantic predictions.
This is the 7th edition of a workshop series that started from landmark recognition alone (CVPR18,CVPR19) and later broadened its scrope to instance-level recognition (ECCV20,ICCV21,ECCV22,ECCV24). This year we are expanding to include the so called personalized (instance-level) generation models.
Join our Instance-level Recognition and Generation workshop at #ICCV2025 with keynote and oral/poster presentations on image object recognition and generation at its finest granularity; each unique object of the physical world forms its own class.
The colloquium at CTU in Prague had 6 great talks and a lot of discussions before, during and after the event. The slides are now shared online. It was the 50th and our administrators surprised us with a huge Czech cake - KolΓ‘Δ. See you in April again!
