BriefGPT.xyz
Dec, 2021
输入级别感应偏置用于3D重建
Input-level Inductive Biases for 3D Reconstruction
HTML
PDF
Wang Yifan, Carl Doersch, Relja Arandjelović, João Carreira, Andrew Zisserman
TL;DR
本文通过将几何归纳偏置直接注入模型作为额外输入，而非使用特殊架构，来解决3D重建问题。研究将摄像机、投射光线和对极几何编码为模型输入，并在多个基准测试上展示具有竞争力的多视图深度估计性能。
Abstract
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
→