Our (slightly updated) Earthvision paper is out now!

CVPR ‘24 was last week, and our lead author Ivica Obadic presented our paper “Contrastive Pretraining for Visual Concept Explanations of Socioeconomic Outcomes” at the Earthvision workshop. In this paper we introduce post-hoc concept presence testing to socio-economic prediction using optical remote sensing. We show that our method can derive both local and global explanations, and we show that our method can be used to derive concept importance across the full spectrum of examples. In this update we have slightly revised the paper by improving the normalisation, and by fixing a bug in the library we used to compute the concept activation vectors. Our V2 paper is available now.