AFLW-19 (The 19 landmark variant of AFLW.)

原始的 AFLW 为每个人脸提供最多 21 个点，但不包括不可见地标的坐标，这给大多数现有基线方法的训练带来了困难。为了进行公平比较，作者手动注释了这些不可见地标的坐标，以便能够训练这些基线方法。新的注释不包括两个耳朵点，因为很难确定隐形耳朵的位置。这导致 AFLW-19 的点数为 19。原始 AFLW 不提供 train-test 分区。 AFLW-19 采用 20,000 张训练图像和 4,386 张测试图像的分区（AFLW-Full）。此外，提出了一个正面子集（AFLW-Frontal），其中所有地标都是可见的（总共 1,165 个图像）。新的 19 点注释文件可在项目页面上找到。

The original AFLW dataset provides up to 21 landmarks per face, but excludes coordinates of invisible landmarks, which creates training challenges for most existing baseline methods. To enable fair comparison, the authors manually annotated the coordinates of these invisible landmarks to allow training of these baseline methods. The revised annotation excludes two ear landmarks, as it is difficult to determine the positions of invisible ears, resulting in AFLW-19 having 19 landmarks per face. The original AFLW does not provide a predefined train-test split. AFLW-19 adopts a split consisting of 20,000 training images and 4,386 test images, referred to as AFLW-Full. Furthermore, a frontal-facing subset named AFLW-Frontal is proposed, where all landmarks are fully visible, containing a total of 1,165 images. The updated 19-landmark annotation files are available on the project webpage.

搜集汇总

背景与挑战

背景概述

AFLW-19是原始AFLW数据集的19个关键点变种，包含20,000张训练图像和4,386张测试图像，特别提供了不可见地标的坐标注释以支持基线方法训练。该数据集还包括一个1,165张图像的正面子集，所有地标均为可见状态。