Dublin, here we come! 🇮🇪 Our paper "Dynamic Scoring with Enhanced Semantics for Training-Free Human-Object Interaction Detection" has been accepted to #ACMMM2025

Thanks to the greatest @massi_manc, @DonkeyShot21, @YimingWang107, @paolorotaphd, and @eliricci_ for their support — couldn’t have done it without you! 🙏🏻 📄 Abstract: https://t.co/7M1SgyXc7w 💻 Code:

We try prompting + reasoning strategies to flip wrong → right and generic → specific. They help, but it’s clear that open-world classification is still very challenging for today’s LMMs.

These distinctions really matter. Most models struggle with fine-grained reasoning, but not always in obvious ways. We dig into why they fail and how to make them better, without retraining.

Spoiler: it’s hard. To evaluate their free-form predictions, we propose metrics that let us classify responses into 4 types: ✅ Correct & specific ("Pug") ✅ Correct but generic ("Dog") ❌ Wrong but specific ("Siamese" vs "Persian") ❌ Wrong & generic ("Plate" vs "Tiramisu")

We tested 13 Large Multimodal Models across 10 datasets — from prototypical to very fine-grained — using natural language instead of fixed label sets. Can LMMs answer: "What type of object is in the image?" …without a predefined list?

What if we stopped treating image classification like a multiple-choice quiz… …and just asked the model: "What’s in this image?" Our paper on open-world classification with LMMs got into #ICCV2025! 🎉🌺 Let’s talk failures, insights, and flipping mistakes 👇

I also got a minimalistic reviewer. Guess the rating

Our T3AL 🦆 poster has been picked by @CVPR daily! Come by and talk to @bliberatori_ @altndrr @paolorotaphd me, and @eliricci_ in the afternoon session!

This is incredible🤯 At test-time, ONE unlabelled video is enough to teach CLIP-like models to localize actions in time despite their lack of prior knowledge of the time domain. More on our @CVPR paper below 👇

Glad to announce that the 1st Workshop on Green Foundation Models (GreenFOMO) is happening @eccvconf Milan, 2024🥳 Can't wait to see what impact our community can make towards a green world 🌏and an inclusive AI💚! Official website + CfP coming soon...Stay tuned😉

A List of Predictions Made in 1924 About 2024 🧵

Cool universal classifier that leverage a knowledge database. You take a pic and get a class, without having defined the classes in advance.

Card design by @taapstudio, check out their Instagram profile ( https://t.co/D7x8xDpQI6)! 🔥

Do you know the names of these animals? 🦆🐑🐖 Tomorrow morning at #NeurIPS (poster #201), we will gift these cards and more to try the demo of our method CaSED! 🔥 Try it now ( https://t.co/GgDqhgijaq) and classify images with CLIP without a predefined list of class names! 🚀

Currently at #NeurIPS for our work "Vocabulary-free Image Classification"! 🚀 We will present the poster this Wednesday morning, don't miss it! 👀

I'll present our work on vocabulary-free image classification in a few minutes at the @SFScon. Streaming here: https://t.co/Hx1kWYgiQ7 @cimec_unitrento @UniTrento_DISI

🌟 Job Opportunity! 🌟 Join my research team at @UniTrento_DISI as a Postdoctoral Researcher in Computer Vision 🖥️ working on a project for creating Bias-Free models for Visual Recognition 📸 If you are interested contact me by email! 🚀 #AI #ComputerVision #ResearchJobs

So happy to share that our paper got accepted at #NeurIPS 2023! 🚀 Preprint: https://t.co/KD9vC82WV6 Code: https://t.co/37srbGw4dh Demo and model on @huggingface 👇 Thanks again to the greatest co-authors @DonkeyShot21 @massi_manc @paolorotaphd @YimingWang107 and @eliricci_! 🤗

Paper is out! 👇👇 #ICCV2023