egoPPG: Heart Rate Estimation from Eye-Tracking Cameras in Egocentric Systems to Benefit Downstream Vision Tasks. Talk in: Project Aria Research Tutorial ICCV'25 poster: Tuesday arXiv: https://t.co/WBkuTGozV6 code+dataset: https://t.co/aZm1a7f9ZH project:

PulseFormer fuses spatial & temporal attention via IMU-video cross-attention to extract HR from stable regions and down-weight motion-afflicted frames. PulseFormer is 24% better than SOTA rPPG—and our HR annotations on EgoExo4D boost the proficiency estimation benchmark by 14%!

Aria 2 now has a heart-rate sensor. But fret not, our novel method PulseFormer can retrofit egocentric video datasets from Aria 1 with continuous HR. #ICCV2025 We also propose egoPPG, a new task to estimate HR directly from eye-tracking cameras in AR/VR headsets. @ICCVConference

📸 Are you adjusting photos? Forget rigid filters—just use Meta-PO! Our #UIST2025 paper combines meta-learning and preference optimization, helping you discover personalized filters simply by selecting your favorite design 5–6 times. 👉 Read more: https://t.co/pUPy2a3LKP

reference: Yao Song, Christoph Gebhardt, Yi-Chi Liao, and Christian Holz. Preference-Guided Multi-Objective UI Adaptation. ACM UIST 2025. @ACMUIST paper: https://t.co/UHTztGUhvp full video: https://t.co/6pql1iLUEA project page: https://t.co/mQn0hwbo5A

Our approach * infers user preferences from their UI adjustments * groups objectives into priority levels based on these preferences * steers Pareto search with lexicographic multi-objective optimization This reduces manual tweaking and produces layouts that match user intent.

Adaptive UIs in Mixed Reality are tricky: Pareto-based optimization gives many optimal layouts, but which one fits best? Introducing preference-guided multi-objective optimization, a method that adapts UIs to what users want via demonstration and semantic alignment. #UIST2025

reference: Andela Ilic, J. Jiang, P. Streli, X. Liu, C. Holz. Human Motion Capture from Loose and Sparse Inertial Sensors with Garment-aware Diffusion Models. Int. Joint Conf. on Artificial Intelligence 2025 @IJCAIconf arXiv: https://t.co/XZr6v0M1T5 more:

We train our diffusion models on – simulated IMU signals – synthetic loose-wear signals via diffusion – real-world recordings with motion artifacts With garment parameters (fit, looseness, body shape), our method stays robust to shifting sensors while capturing accurate motion.

Making motion capture a lot more practical: Garment Inertial Poser estimates full-body motion from IMUs that are loosely attached to everyday clothing—no cameras, no tight straps. #IJCAI2025 Talk is today in the session "Diffusion models". Andela Ilic @cs_jiaxi_jiang @paulstreli

reference: Yi-Chi Liao, Paul Streli, Zhipeng Li, Christoph Gebhardt, Christian Holz. Continual Human-in-the-Loop Optimization. ACM CHI 2025. @acm_chi paper: https://t.co/kkRhAE9bFB arXiv: https://t.co/dMXcNq5L8v project page:

At its core is a scalable Bayesian neural network trained via generative memory replay across past users. It avoids forgetting and gets better with use. Benefits: ✨ users get personalized UIs. 🛠️ designers need to tweak controls less. 💡 developers can build adaptive systems.

Tired of configuring XR UIs, so that controls are well positioned, feel right, & match your preferences? Introducing Continual Human-in-the-Loop Optimization, a framework that learns what works across users and personalizes UIs for new users faster every time. #CHI2025 @CSatETH

Berken Utku Demirel and Christian Holz. Shifting the Paradigm: A Diffeomorphism Between Time Series Data Manifolds for Achieving Shift-Invariancy in Deep Learning. ICLR 2025 @iclr_conf arXiv: https://t.co/WH3QR7r3A4 code: https://t.co/nLUgGzj8xn project:

Our deep learning method ✅ removes the limited shift assumption by using the continuous translation group ✅ boosts performance by 10–15% over SOTA across 6 diverse tasks ✅ works across domains—audio, IMUs, biosignals, etc. ✅ is drop-in compatible with existing architectures

Come see our paper on shift invariance in ML at #ICLR25 today. Our method makes any model shift-invariant—without changing architectures or small shifts. Shift-invariancy in temporal data is a key challenge for robust models in health, behavior analysis, and beyond. @CSatETH

We propose a new theory of affordance grounded in Computational Rationality: Affordances aren’t static—they're learned, refined, and contextual. Our model explains how we learn, develop, & perceive digital, social, and false affordances. #IUI2025 #HCI #AI https://t.co/gdDy69TVHS

🤔 Why do we perceive some real-world actions that aren't actually possible—and miss the ones that are? In our @ACMIUI paper, we frame affordance as users' inference over confidence of execution and predicted utility of the action—unlike Gibson's "information pickup." @CSatETH

B. Demirel, A. Dogan, J. Rossie, M. Moebus, C. Holz Beyond Subjectivity: Continuous Cybersickness Detection Using EEG-based Multitaper Spectrum Estimation IEEE TVCG '25 @ieee_tvcg paper: https://t.co/5FVOv8PEt3 code+dataset: https://t.co/ZDpyVPmXZt more:

Using EEG & head motion as input, our multitaper-based approach extracts effective representations from these multimodal inputs. Two encoders then derive modality-specific features from the spectral density, fuse them, and continuously estimate a user's sickness levels.