FLAT

场景运动、多次反射和传感器噪声会在飞行时间相机执行的深度重建中引入伪影。我们提出了一种两阶段的深度学习方法来同时处理所有这些伪像来源。我们还介绍了 FLAT，这是一个包含 2000 个 ToF 测量的合成数据集，可以捕获所有这些非理想情况，并可用于模拟不同的硬件。使用 Kinect 相机作为基线，与最先进的方法相比，我们在模拟和真实数据上展示了改进的重建误差。

Scene motion, multipath reflections, and sensor noise introduce artifacts into the depth reconstruction performed by time-of-flight (ToF) cameras. We propose a two-stage deep learning framework to simultaneously mitigate all these sources of artifacts. We additionally introduce FLAT, a synthetic dataset encompassing 2000 ToF measurements that captures all these non-ideal scenarios and enables simulation of diverse hardware configurations. Using the Kinect camera as our baseline, we demonstrate reduced reconstruction errors on both simulated and real-world data when compared against state-of-the-art methods.

搜集汇总

背景与挑战

背景概述

FLAT是一个包含2000个ToF测量的合成数据集，用于处理飞行时间相机深度重建中的伪影问题，支持模拟不同硬件条件，由哈佛大学于2018年发布。