交叉姿势 LFW 数据库 (CP-LFW) 人脸识别中交叉姿势鲁棒性的数据集

欢迎来到Cross-Pose LFW (CPLFW)数据库，它是Labeled Faces in the Wild (LFW)的翻新版，是无约束人脸验证的事实标准测试平台。 野生标签脸（LFW）数据库已被广泛用作无约束人脸验证的基准，由于大数据驱动的机器学习方法，该数据库的性能几乎接近100%。然而，我们认为这一准确性可能过于乐观。除了不同的光照、遮挡和表情，交叉姿势的人脸是人脸识别的另一个挑战，但LFW并没有对其给予太多的关注。因此，我们构建了一个跨姿势的LFW（CPLFW），它特意搜索并选择了3000个有姿势差异的正脸对，以增加类内差异的姿势变化。同时，还选择了具有相同性别和种族的阴性对，以减少阳性/阴性对之间属性差异的影响。我们在新数据库中评估了几种深度学习方法。与LFW上的准确率相比，CPLFW上的准确率下降了大约15%-20%。构建CPLFW基准的背后有三个动机，如下。 1、建立一个相对更难的数据库来评估真实世界人脸验证的性能，这样就可以充分证明几种人脸验证方法的有效性。 2、继续深入研究LFW，更真实地考虑姿态的类内变化，并促进无约束情况下的跨姿态人脸验证研究。CPLFW的挑战在于强调姿势差异，以进一步扩大类内差异。此外，还特意选择了负数对以避免不同的性别或种族。CPLFW同时考虑了大的类内方差和小的类间方差。 3、保持数据大小，提供 "相同/不同 "基准的人脸验证协议和LFW中的相同身份，所以人们可以很容易地应用CPLFW来评估人脸验证的性能。 We dedicate to maintain the protocols, dataset size, and the identities in each fold of LFW database in order to encourage fair and meaningful comparisons. You can find more information about standard LFW protocol in Labeled Faces in the Wild (LFW). We expect CPLFW could promote algorithms to make reliable verification judgement, and close the large gap between the reported performance on benchmarks and performance on real world tasks.

Welcome to the Cross-Pose LFW (CPLFW) database, a revised version of the Labeled Faces in the Wild (LFW) database and the de facto standard benchmark platform for unconstrained face verification. The Labeled Faces in the Wild (LFW) database has been widely used as a benchmark for unconstrained face verification, and with the advancement of big data-driven machine learning methods, its reported performance has nearly reached 100% accuracy. However, we argue that this accuracy may be overly optimistic. In addition to variations in illumination, occlusion, and expression, cross-pose faces pose another major challenge for face recognition, but the original LFW database has paid little attention to this aspect. Therefore, we constructed the Cross-Pose LFW (CPLFW) database, which specifically searched and selected 3000 positive face pairs with pose differences to increase the intra-class variations caused by pose changes. Meanwhile, we also selected negative pairs with matching gender and race to mitigate the impact of attribute differences between positive and negative pairs. We evaluated several deep learning methods on this new database. Compared with the accuracy reported on LFW, the accuracy on CPLFW decreased by approximately 15% to 20%. There are three core motivations behind the construction of the CPLFW benchmark, as follows:
1. Establish a relatively more challenging database to evaluate the performance of real-world face verification, so as to fully demonstrate the effectiveness of multiple face verification methods.
2. Further conduct in-depth research on LFW, more realistically consider the intra-class variations of pose, and promote the research on cross-pose face verification in unconstrained scenarios. The core challenge of CPLFW lies in emphasizing pose differences to further expand intra-class variations. In addition, negative pairs are specifically selected to avoid discrepancies in gender or race. CPLFW simultaneously considers large intra-class variance and small inter-class variance.
3. Maintain the dataset size, provide the same "same/different" face verification protocol as in LFW with consistent identities, so that researchers can easily apply CPLFW to evaluate the performance of face verification.
We are committed to maintaining the protocols, dataset size, and identities in each fold of the LFW database to encourage fair and meaningful comparisons. You can find more information about the standard LFW protocol in the original Labeled Faces in the Wild (LFW) paper. We expect that CPLFW can promote the development of algorithms to make reliable verification judgments, and narrow the large gap between the reported performance on benchmarks and the actual performance in real-world tasks.

搜集汇总

背景与挑战

背景概述

CP-LFW是一个专注于跨姿态人脸识别的数据库，通过精心挑选3000个有显著姿势差异的正脸对和相同性别种族的阴性对，来增强类内差异和减少属性差异的影响。与LFW相比，CP-LFW在多个先进深度学习方法上的识别准确率下降了15%-20%，为研究真实世界中的人脸验证提供了更具挑战性的基准。