PhotoMatte85

我们介绍了一种实时、高分辨率的背景替换技术，该技术在 4K 分辨率下以 30fps 运行，在现代 GPU 上以 60fps 运行高清。我们的技术基于背景遮罩，其中额外的背景帧被捕获并用于恢复 alpha 遮罩和前景层。主要挑战是计算高质量的 alpha 遮罩，保留股线级头发细节，同时实时处理高分辨率图像。为了实现这一目标，我们采用了两个神经网络；一个基础网络计算一个低分辨率结果，该结果由第二个网络在选择性补丁上以高分辨率运行。我们介绍了两个大型视频和图像抠图数据集：VideoMatte240K 和 PhotoMatte13K/85。与之前最先进的背景消光技术相比，我们的方法产生了更高质量的结果，同时显着提高了速度和分辨率。

We introduce a real-time, high-resolution background replacement technique that runs at 30 fps for 4K resolution and 60 fps for high-definition (HD) content on modern GPUs. Our approach is based on background matting, where additional background frames are captured and utilized to recover alpha mattes and foreground layers. The primary challenge lies in computing high-quality alpha mattes while retaining strand-level hair details and processing high-resolution images in real time. To address this challenge, we employ two neural networks: a base network that generates a low-resolution result, which is then refined by a second network operating on selective patches at high resolution. We present two large-scale video and image matting datasets: VideoMatte240K and PhotoMatte13K/85. Compared with prior state-of-the-art background matting methods, our approach delivers higher-quality results while substantially improving both speed and resolution.