Great to have a workshop on object-centric representations again after a dry spell last year!

World models are key for developing adaptive agents. In our #ICLR2024 spotlight we present THICK: an algorithm to learn hierarchical world models with versatile temporal abstractions. And we show how they can enhance model-based RL or planning… 📜 openreview.net/forum?id=TjCDN… 1/🧵

I am in Vienna this week, presenting our #ICLR2024 spotlight: the Dynamic Scene Transformer, a video model that learns a structured representation of camera, motion and content. Come by Fri 10:45, poster 100. Or reach out if you want to chat!

Happy to share that our paper - On the Transfer of Object-Centric Representation Learning - is accepted at ICLR 2025! ICLR 2026 Work with - Max Seitzer Andrii Zadaianchuk 🇺🇦 Anirudh Goyal Michael C. Mozer Yoshua Bengio Georg Martius 🔗 rw-ocrl.github.io/ftdinosaur-pap… 📜 openreview.net/forum?id=bSq0X…

🔥Happy to share that our paper on controllable object-centric representations (CTRL-O) has been accepted to #CVPR2025 #CVPR2025 w/ Andrii Zadaianchuk 🇺🇦 Rabiul Awal Max Seitzer Efstratios Gavves Aishwarya Agrawal 🔜 Updated paper with downstream experiments and detailed thread coming soon!

I will be presenting our paper “On the Transfer of Object-Centric Representation Learning” in Hall 3 + Hall 2B #106 on Saturday 26th April from 10-12:30. See below thread for more details 👇

🧵𝗪𝗮𝗻𝘁 𝘆𝗼𝘂𝗿 𝗺𝗼𝗱𝗲𝗹 𝘁𝗼 𝗽𝘂𝗹𝗹 𝗼𝘂𝘁 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗼𝗯𝗷𝗲𝗰𝘁 𝗿𝗲𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗳𝗿𝗼𝗺 𝗮 𝘀𝗰𝗲𝗻𝗲 𝗼𝗻 𝗱𝗲𝗺𝗮𝗻𝗱-𝗷𝘂𝘀𝘁 𝗯𝘆 𝗱𝗲𝘀𝗰𝗿𝗶𝗯𝗶𝗻𝗴 𝗶𝘁? Meet CTRL-O, our #CVPR2025 method for language-driven object-centric representations.👇

Thrilled our SlotContrast paper is an oral at CVPR25! 🎉 SotA object-centric video representations, learned *fully unsupervised* by maximizing temporal consistency — makes DINO features instance-aware as a bonus 🦕. Amazing job, Anna! Check out: slotcontrast.github.io

Introducing SlotContrast, state-of-the-art object-centric representation learning for real-world videos and an oral at CVPR! Paper thread 👇