Nikita Araslanov
Nikita Araslanov is a visiting researcher at the Visual Geometry Group (VGG), University of Oxford. Previously, he was a postdoctoral scientist at TUM (2022–2026) and a visiting faculty researcher at Google (2024–2025). His research focuses on bridging semantic and 3D visual inference using video data. He received his PhD in Computer Science from TU Darmstadt in 2022 and completed his master’s degree in Computer Science at the University of Bonn.
Project
Despite rapid progress in visual learning, we still lack understanding about what structural properties make some pretrained models exceptionally useful in practice while other models see limited adoption. This project aims to link the observed practical success to a distinctive symmetry profile of the learned representation. Using curated real-world videos with point trajectories, we will construct track-based kernels and derive a small set of mathematically grounded, low-overhead, and interpretable metrics for three transformation classes: camera motion, independent object motion, and interactions.
The central hypothesis is that useful models exhibit selective invariance to nuisance changes, structured equivariance under object motion, and clear symmetry-breaking signatures at interactions. Drawing on causal representation learning, object-centric vision, and kernel/spectral analysis, the project will turn these ideas into precise, testable metrics. The emphasis is not on model training or large-scale benchmarking, but on understanding off-the-shelf models through carefully designed hypotheses and normalized metrics. The workflow is laptop-friendly, and the annotated data will be prepared in advance.
