Analysis of Visual Foundation Models with Applications to Video Coding for Machines

Beschreibung

Visual Foundation models (VFM), like DINOv3, V-JEPA,… are widely employed for diverse computer vision tasks.
Unlike traditional computer vision pipelines, VFMs learn rich, transferable representations from vast amounts of data, enabling strong generalization across tasks with little to no finetuning.
This research shall explore the usage of VFM models in machine-to-machine communication of visual data.

Mögliche Themen

• Cross-architecture feature translation
• Intra frame feature prediction
• Inter frame feature prediction (forecasting)
• VFM feature compression

Voraussetzungen

Vorausgesetzt werden Erfahrungen in der Programmierung mit Python, Deep Learning (PyTorch/TF) und Bild- und Videokompression