Computer Vision and Machine Learning Department (D2) at the Max Planck Institute for Informatics

Congratulations to Yonggang Jiang on receiving one of the 2025 Google PhD Fellowships in the area of “Algorithms and Optimization.” 🥳 🎓 The Google PhD Fellowships recognize and support outstanding doctoral candidates conducting innovative research in computer science and related fields. Yonggang is broadly interested in the design and analysis of algorithms. His current research focuses on graph algorithms, with the goal of understanding their theoretical limits across various settings, including parallel and distributed computing. Yonggang joined the Max Planck Institute for Informatics in the Algorithms and Complexity department of Danupon Nanongkai in 2023. ↪️ For more info, visit his homepage: https://lnkd.in/eetM-Zjb

Two new Emmy Noether research groups established at the MPI for Informatics, led by Dr. Jan Eric Lenssen and Dr. Karol Węgrzycki. Together, they will receive approximately 3.7 million euros in funding from the Deutsche Forschungsgemeinschaft (DFG) - German Research Foundation to set up their groups in Saarbrücken. Their research broadly focuses on algorithms and machine learning. 🎓 📌 Dr. Jan Eric Lenssen’s group, Spatial Modeling and Reasoning, will explore how AI systems can extract more infromation from sparse visual data. To this end, he will work on efficient spatial representations and machine learning methods that perform spatial reasoning across these representations. 📌 Dr. Karol Węgrzycki’s research group will investigate packing problems: a class of optimization challenges that appear in many practical situations ranging from cargo loading to material cutting, and how insights from different areas of theoretical computer science can be combined to solve them more efficiently. Both researchers are now part of the prestigious Emmy Noether Programme, which supports outstanding early-career scientists in leading independent groups for up to six years, qualifying them toward a university professorship. Join us in congratulating Jan and Karol on this remarkable achievement! 👏 ↪️ Learn more: https://lnkd.in/eYAVTuzx

🚀 Thrilled to announce RefAM, our latest work with Enis Simsar, a training-free method showing that diffusion models already contain the signals needed for referring segmentation within their own attention maps. Traditional referring segmentation methods rely on large finetuned networks and CLIP. In RefAM, we show that diffusion transformers already encode strong localization cues if we read their attention properly. Some text tokens (like “the”, “is”) act as Global Attention Sinks (GAS) that diffuse focus across the image. We introduce Attention Magnets, harmless extra tokens that absorb this noise and can be filtered out to yield clean, localized attention maps, no training required. RefAM achieves +2.5 mIoU over HybridGL on RefCOCO/+/g and performs strongly on Ref-DAVIS17, Ref-YouTube-VOS, and MeViS, all in a zero-shot setting. 📄 arxiv.org/abs/2509.22650 💥 refam-diffusion.github.io 💻 github.com/Annusha/refam By Anna Kukleva, Enis Simsar, Alessio Tonioni, Muhammad Ferjad Naeem, Federico Tombari, Jan Eric Lenssen, and Bernt Schiele #DiffusionModels #ComputerVision #Segmentation

ICCV 2025 🌺 Aloha from Hawaii! Our group at MPI-INF (Dept. D2) is presenting four papers this year — including one Highlight. Brief overviews + links below 👇 1) Do It Yourself: Learning Semantic Correspondence from Pseudo-Labels (DIY-SC) A lightweight adapter on DINOv2 / SD+DINOv2 that reaches SOTA on SPair-71k without annotated keypoints. Authors: Olaf Dünkel, Thomas Wimmer, Christian Theobalt, Christian Rupprecht, Adam Kortylewski 🔗 Project: https://lnkd.in/e85rG4WM 🔗 Related post: https://lnkd.in/eQzSudPM 2) AIM (Highlight): Amending Inherent Interpretability via Self-Supervised Masking A simple, effective approach that encourages models to use genuine features over spurious ones—no extra annotations (only image labels). Authors: Eyad Alshami, Shashank Agnihotri, Bernt Schiele, Margret Keuper 🔗 Project: https://lnkd.in/e7mGA5HN 3) VITAL: More Understandable Feature Visualization through Distribution Alignment & Relevant Information Flow Reduces repetitive artifacts, improves interpretability, and scales to modern large-scale architectures. Authors: Ada Görgün, Bernt Schiele, Jonas Fischer 🔗 Project: https://lnkd.in/eH2cVxQ9 🔗 Post: https://lnkd.in/e8qRhitA 4) MVGBench: A Comprehensive Benchmark for Multi-view Generation Models A fair, comparable suite for object reconstruction with emphasis on 3D consistency & image quality. Authors: Xianghui Xie, Jan Eric Lenssen, Gerard Pons-Moll 🔗 Project: https://lnkd.in/e3M46-2p 🔗 X thread: https://lnkd.in/ehgNHg_B If you’re at ICCV, come say hi! 🤝 #ICCV2025 #ComputerVision #DeepLearning #Interpretability #RepresentationLearning #3D #Benchmarking #MPIINF

Attending #ICCV2025 🏝️🌺? Don't miss our full-day Explainable Computer Vision (eXCV) Workshop this Sunday in Ballroom A. We feature an outstanding lineup of speakers, insightful panel discussions, and papers spanning three exciting tracks! Checkout the program and further details: https://lnkd.in/eja3N7eB Invited Speakers: David Bau, Deepti Ghadiyaram, Hila Chefer, Sharon Li, Thomas Fel, Viorica Patraucean Organizers: Sukrut Rao, Robin Hesse, Quentin Bouniot, Sweta Mahajan, Amin Parchami, Jayneel Parekh, Simone Schaub-Meyer, Florence d'Alche-Buc, Zeynep Akata, Stefan Roth, Bernt Schiele #XAI #Explainability #Interpretability #Workshop #excv #ICCV #ICCV2025

Heading to #ICCV2025? Don’t miss #CV4DC — Oct 19, Honolulu 🌺 An in-person workshop spotlighting research for/with developing regions. Keynotes: Prof. Lorenzo Torresani, Asst. Prof. Georgia Gkioxari, Prof. Sai-Kit Yeung and Dr. Ronnachai Jaroensri. See accepted papers & posters here: https://lnkd.in/eTpYSCc2 #AI4Good #ComputerVision

🚀 Call for Questions! 🚀 Got questions about the state of XAI in Computer Vision? Is there a challenge you want to spotlight, a discussion you would like to ignite, or a change you hope to see in the field? We are inviting the community and the stakeholders to submit questions, which will be discussed with our experts at the workshop! 🎤 💡 Submit your questions: https://lnkd.in/dx46Rsp3 Workshop Website: https://lnkd.in/dAtcmp-G Organizers: Sukrut Rao, Robin Hesse, Quentin Bouniot, Sweta Mahajan, Amin Parchami, Jayneel Parekh, Simone Schaub-Meyer, Florence d'Alche-Buc, Zeynep Akata, Stefan Roth, Bernt Schiele #XAI #eXCV #foundationalmodels #LanguageExplanation #computerVision #NewExplanationMethod #Scale

⏳Still need to wait for your last experiment results? 📣 We're pleased to announce that the deadline for non-proceeding track #CV4DC at @iccv.bsky.social has been extended to August 15, 2025 Looking forward to your submissions! cv4dc.github.io/2025/

🩻 Can AI truly offer clinically reliable explanations for chest X-ray diagnosis? We explore this question with our latest work: Faithful, Interpretable Chest X-ray Diagnosis with Anti-Aliased B-cos Networks ➡️ https://lnkd.in/eUnpgg8h 🔍 What is new? ✅ We introduce anti-aliasing techniques (FLCPooling and BlurPool) to eliminate visual artifacts in explanation maps ✅ Our method preserves strong predictive performance and produces saliency maps that are both faithful and artifact-free 🏥 These improvements bring us a step closer to AI systems that clinicians can interpret, trust, and verify. 🧠 If you work at the intersection of medical imaging, explainability, and deep learning, we would love your thoughts and feedback! This work is by Marcel Kleinmann, supervised by Margret Keuper (Data and Web Science Group at University of Mannheim, Computer Vision and Machine Learning Department (D2) at the Max Planck Institute for Informatics) #MedicalImaging #ExplainableAI #DeepLearning #ChestXray #InterpretableModels #XAI #AIForHealth