ICCV21Deep

该数据集包含96个高质量的场景，这些场景显示了户外场景中的各种视觉效果和人类互动。我们开发了一种新算法Deep 3D Mask Volume，该算法可以从静态摄像机捕获的动态场景的双目视频中进行时间稳定的视图外推。我们的算法通过使用3D遮罩体积识别容易出错的区域来解决解遮挡的时间不一致，并将其替换为整个视频中观察到的静态背景。与简单的2D遮罩相反，我们的方法可以在3D空间中进行操作，与逐帧静态视图合成方法或使用2D遮罩的方法相比，我们展示了更好的时间稳定性。生成的视图合成视频显示最少的闪烁伪影，并允许更大的平移运动。

This dataset comprises 96 high-quality scenes depicting various visual effects and human interactions in outdoor environments. We have developed a novel algorithm named Deep 3D Mask Volume, which enables temporally stable view extrapolation from binocular videos of dynamic scenes captured by static cameras. Our algorithm addresses temporal inconsistency in disocclusion by identifying error-prone regions using 3D mask volumes, and replaces these regions with static backgrounds observed across the entire video. In contrast to simple 2D masks, our method operates in 3D space, and exhibits superior temporal stability compared to per-frame static view synthesis methods or those using 2D masks. The resulting view synthesis videos exhibit minimal flickering artifacts and enable larger translational movements.

搜集汇总

背景与挑战

背景概述

ICCV21Deep数据集包含96个高质量户外场景，用于研究视觉效果和人类互动。它采用Deep 3D Mask Volume算法，从双目视频实现时间稳定的视图外推，通过3D遮罩体积减少闪烁伪影，提升平移运动稳定性。该数据集由Facebook Reality Labs和加州大学伯克利分校于2022年发布。