300-W
收藏资源简介:
300-W 是一个人脸数据集,由 300 张室内和 300 张室外野外图像组成。它涵盖了身份、表情、照明条件、姿势、遮挡和面部大小的大量变化。这些图片是通过查询“派对”、“会议”、“抗议”、“足球”和“名人”等从 google.com 下载的。与其他野外数据集相比,300-W 数据库包含更大比例的部分遮挡图像,并且涵盖的表情比常见的“中性”或“微笑”(例如“惊喜”或“尖叫”)更多.使用半自动方法用 68 点标记对图像进行注释。数据库中的图像经过精心挑选,因此它们代表了在完全不受约束的条件下具有挑战性但自然的人脸实例的特征样本。因此,在 300-W 数据库上实现准确性能的方法可以在大多数实际情况下展示相同的准确度。数据库中的许多图像包含不止一张带注释的人脸(293 张带有 1 张人脸的图像,53 张带有 2 张人脸的图像和 53 张带有 [3, 7] 人脸的图像)。因此,该数据库包含 600 个带注释的人脸实例,但包含 399 个独特的图像。最后,有各种各样的脸型。具体来说,49.3% 的人脸大小在 [48.6k, 2.0M] 范围内,整体平均大小为 85k(约 292 × 292)像素。
300-W is a facial landmark dataset consisting of 300 indoor and 300 outdoor in-the-wild images. It covers extensive variations in identity, expression, lighting conditions, pose, occlusion and facial size. These images were downloaded from google.com via queries using keywords such as "party", "meeting", "protest", "soccer" and "celebrity". Compared with other in-the-wild datasets, the 300-W database contains a larger proportion of partially occluded images, and covers a wider range of expressions beyond common "neutral" or "smiling" ones, such as "surprised" or "screaming". The images were annotated with 68-point landmarks using a semi-automatic method. The images in the database were carefully selected to represent a characteristic sample of challenging yet natural facial instances under fully unconstrained conditions. Therefore, methods achieving accurate performance on the 300-W database can demonstrate comparable accuracy in most real-world scenarios. Many images in the database contain more than one annotated face: 293 images with 1 face, 53 images with 2 faces, and 53 images with [3, 7] faces. As a result, the database contains 600 annotated facial instances but only 399 unique images. Finally, there is a wide variety of face shapes. Specifically, 49.3% of the faces have a size within the range of [48.6k, 2.0M] pixels, with an overall average size of 85k (approximately 292 × 292) pixels.

- 300-W数据集首次发表,由Sagonas等人提出,旨在解决面部特征点定位问题。
- 300-W数据集被广泛应用于面部特征点定位算法的研究和评估,成为该领域的重要基准。
- 随着深度学习技术的发展,300-W数据集开始被用于训练和测试基于深度神经网络的面部特征点定位模型。
- 300-W数据集的扩展版本300-W-LP发布,增加了更多的合成数据,以提高模型的泛化能力。
- 300-W数据集及其扩展版本继续被广泛用于面部特征点定位和相关任务的研究,推动了该领域的技术进步。
- 1300 Faces In-The-Wild Challenge: Database and ResultsImperial College London · 2016年
- 2Deep Convolutional Network Cascade for Facial Point DetectionUniversity of Science and Technology of China · 2013年
- 3Wing Loss for Robust Facial Landmark Localisation with Convolutional Neural NetworksUniversity of Adelaide · 2018年
- 4A Deep Regression Architecture with Two-Stage Re-initialization for High Performance Facial Landmark DetectionUniversity of California, Irvine · 2017年
- 5Facial Landmark Detection by Deep Multi-task LearningUniversity of Science and Technology of China · 2014年



