Determine the nature of what DINOv2 perceives

Determine the nature of what DINOv2 perceives as reflected in its internal visual representations, by characterizing the content and structure of the features encoded by DINOv2.

Background

DINOv2 achieves strong performance across numerous vision tasks without explicit labels, suggesting rich internal representations. However, the specific content and organization of these representations are not directly observable and have not been fully characterized.

This uncertainty motivates the paper’s investigation using sparse autoencoders and, later, the proposed Minkowski Representation Hypothesis to probe what features are present and how they may be geometrically organized.