Input-level Inductive Biases for 3D Reconstruction
This approach addresses the need for efficient and adaptable 3D vision models for researchers and practitioners, though it is incremental as it builds on existing methods by modifying inputs rather than architectures.
The paper tackled 3D reconstruction by injecting geometrical inductive biases as extra inputs to a domain-agnostic architecture, enabling the use of general models like Perceivers without architectural changes while maintaining data efficiency, and demonstrated competitive multi-view depth estimation performance on multiple benchmarks.
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 tackle 3D 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 approach makes it possible to apply existing general models, such as Perceivers, on this rich domain, without the need for architectural changes, while simultaneously maintaining data efficiency of bespoke models. In particular we study how to encode cameras, projective ray incidence and epipolar geometry as model inputs, and demonstrate competitive multi-view depth estimation performance on multiple benchmarks.