[论文] OmniRay: Camera-Agnostic Monocular 3D Reconstruction with Composite Ra...

论文概要

研究领域: CV 作者: Namitha Guruprasad, Abhay Yadav, Cheng Peng, Rama Chellappa, Anand Bhattad 发布时间: 2026-10-22 arXiv: 2610.31645

中文摘要

单目模型可以一次前向传播从单张图像恢复 3D 几何，但大多数假设针孔相机模型，在鱼眼和全景镜头产生的广角图像上会产生畸变几何。相机感知模型通过预测逐像素光线场消除了这一假设，但它们用单一固定基来表示不同的相机几何。我们认为相机几何具有层次化的复杂度。OmniRay 的前馈模型 Ray Module 由三个组件构成：捕获径向对称结构的单调径向基、表示平滑偏差的球谐残差，以及捕捉剩余局部变化自由残差。这些组件作为球面上的连续旋转进行组合，使表示能容纳多样的相机几何。给定单张图像，OmniRay 为每个像素预测光线方向和 up-to-scale 的径向距离，从中恢复 3D 点云，推理时无需任何相机信息。跨不同相机配置的广泛实验表明，OmniRay 树立了相机无关单目 3D 重建的新最先进水平。

原文摘要

Monocular models recover 3D geometry from a single image in one pass, but most assume a pinhole camera and produce distorted geometry on wide-angle imagery. Camera-aware models remove this assumption by predicting a per-pixel ray field, but they represent different camera geometries using a single fixed basis. We present OmniRay, whose Ray Module decomposes into three components: a monotone radial base, a spherical-harmonic residual, and a free-form residual, composed as successive rotations on the sphere. Given a single image, OmniRay predicts a ray direction and up-to-scale radial distance for every pixel without needing camera information at inference time, establishing a new state-of-the-art in camera-agnostic monocular 3D reconstruction.

#论文 #arXiv #CV #小凯