Mill 19

我们针对多个不同的数据集评估Mega-NeRF。我们的Mill 19数据集由我们在一个前工业园区附近第一手记录的两个场景组成。Mill 19-Building由在工业建筑周围的大型500 250平方米区域中以网格图案捕获的镜头组成。19号磨坊-瓦砾覆盖了附近一个充满碎片的建筑区域，我们在其中放置了伪装成幸存者的人体模型。我们还针对两个公开可用的集合来测量Mega-NeRF-Quad 6k数据集，在Cornell Universty Arts Quad中收集的大型运动结构数据集以及UrbanScene3D中的几个场景，其中包含高分辨率的无人机图像大规模城市环境。我们使用PixSFM完善了Mill 19和UrbanScene3D数据集中的初始GPS衍生相机姿势以及Quad 6k数据集中提供的估计值。我们使用预先训练的语义分割模型来产生Quad 6k数据集中常见可移动对象的掩码，并在训练过程中忽略掩码像素。

We evaluate Mega-NeRF across multiple distinct datasets. Our Mill 19 dataset consists of two scenes captured firsthand near a former industrial park. Mill 19-Building comprises shots captured in a grid pattern over a large 500×250 square meter area surrounding an industrial building. Mill 19-Rubble covers a nearby debris-filled construction area, where we placed mannequins disguised as survivors. We also evaluate Mega-NeRF on two publicly available collections: the Quad 6k Dataset, a large structure-from-motion dataset collected in the Cornell University Arts Quad, and several scenes from UrbanScene3D, which contains high-resolution drone imagery of large-scale urban environments. We used PixSFM to refine the initial GPS-derived camera poses in the Mill 19 and UrbanScene3D datasets, as well as the estimates provided in the Quad 6k dataset. We employed a pre-trained semantic segmentation model to generate masks for common movable objects in the Quad 6k dataset, and ignored masked pixels during training.

搜集汇总

背景与挑战

背景概述

Mill 19是一个用于评估Mega-NeRF的大规模视觉数据集，包含两个真实世界场景：工业建筑区域和碎片区域，总大小为19.3GB。该数据集由卡内基梅隆大学和Argo AI于2022年发布，支持虚拟飞行通过等应用，具有高分辨率图像和精细的相机姿势处理。