Input-level Inductive Biases for 3D Reconstruction

Much of the recent progress in 3D vision has been driven by the development of specialized architectures that incorporate geometrical inductive biases. In this paper we tackle3D reconstruction using a domain agnostic architecture and study how instead to inject the same type of inductive biases directly as extra inputs to the model. This makes it possible to apply existing general models like Perceivers data-efficiently and without architectural changes on this rich domain. In particular we show how to encode epipolar geometry, projective ray incidence and semantic co-occurence statistics as model inputs and demonstrate state-of-the-art depth estimation performance on ScanNet