Probing the Role of Positional Information in Vision-Language Models

Philipp J. Rösch, Jindřich Libovický

In most Vision-Language models (VL), the understanding of the image structure is enabled by injecting the position information (PI) about objects in the image. In our case study of…