V. Belagiannis and W. Kellermann in Stanford University’s World’s Top 2% Scientists

Stanford University’s World’s Top 2% Scientists list is a prestigious ranking of the top 2% of the most-cited scientists worldwide across all disciplines. It is published annually in collaboration with Elsevier and based on data from Scopus. The latest edition was released in October 2026.
We are delighted to announce that Prof. Dr. Vasileios Belagiannis and Prof. Dr.-Ing. Walter Kellermann from our Chair have been included in this renowned ranking.
Congratulations to both!
This recognition is a proof to our group’s dedication to conducting state-of-the-art research that pushes the boundaries of machine learning and signal processing. Our focus on fundamental challenges and innovative solutions is reflected in our recent publications. Below is a selection of our latest work that showcases our commitment to academic excellence:

  1. “Detecting and Mitigating Memorization in Diffusion Models Through Anisotropy of the Log-Probability” Authors: Rohan Asthana, Vasileios Belagiannis
    This work explains why existing memorization detectors for diffusion models break down at low noise levels and introduces an efficient metric, combining the norm and angular alignment of the log-probability, that detects memorization with only two forward passes.
  2. “Forecasting the Past: Gradient-Based Distribution Shift Detection in Trajectory Prediction” Authors: Michele De Vita, Julian Wiederer, Vasileios Belagiannis. This paper presents a self-supervised, post-hoc method that detects distribution shifts in trajectory prediction for automated driving, using the gradient of a loss for forecasting the second half of an observed trajectory from its first half, without affecting the original prediction model.
  3. “Prompt-Guided Image Editing with Masked Logit Nudging in Visual Autoregressive Models” Authors: Amir El-Ghoussani, Marc Hölle, Gustavo Carneiro, Vasileios Belagiannis. This work introduces Masked Logit Nudging, a method for text-guided image editing with visual autoregressive models that changes only the regions relevant to the prompt and leaves the rest of the image intact.
  4. “Towards Real-Time and Adaptable LiDAR Scene Completion” Authors: Azhar Hussian, Martin Vossiek, Vasileios Belagiannis. This publication presents RapidLiDAR, a LiDAR scene completion method with a learned, adaptive initialization that completes a full scene in 0.1 seconds, matching the 10 Hz scan rate of typical automotive LiDAR sensors.

Source:

  1. https://top2percentscientists.com
  2. https://elsevier.digitalcommonsdata.com/datasets/btchxktzyw/9