CASIA-WEBFACE 人脸图片数据集

在大数据和深度卷积神经网络（美国有线电视新闻网）的推动下，人脸识别的性能已与人类相比。使用私有的大规模训练数据集，若干组在LFW上实现非常高的性能，即97%到99%。虽然有许多开源的美国有线电视新闻网的实现，没有大规模的面部数据集是公开可用的。人脸识别领域的研究现状是数据比算法更重要。为了解决这个问题，我们提出了一种半自动的方式来收集来自互联网的人脸图像，并建立一个大型数据集包含10575个主题和494414个图像，称为CASIA WebFACTS。据我们所知，该数据集的大小在文献中排名第二，仅比脸谱网（SCF）的私有数据集小。我们鼓励在这个数据集上的数据消耗方法训练和LFW上的报告性能。

Driven by big data and deep convolutional neural networks (CNNs), the performance of face recognition has reached parity with that of human beings. Using private large-scale training datasets, several research teams have achieved extremely high performance on the LFW benchmark, ranging from 97% to 99%. Although there are many open-source implementations of CNNs, no large-scale facial datasets are publicly available. The current state of research in the field of face recognition is that data is more critical than algorithms. To address this issue, we propose a semi-automatic method to collect face images from the Internet and construct a large-scale dataset containing 10,575 subjects and 494,414 images, named CASIA WebFACTS. To the best of our knowledge, this dataset ranks second in scale across existing academic literature, only trailing the private dataset of Facebook (SCF). We encourage researchers to train data-hungry models on this dataset and report their performance on the LFW benchmark.

搜集汇总

背景与挑战

背景概述

CASIA-WEBFACE是一个大型公开人脸图片数据集，包含10575个主题和494414个图像，专为人脸识别研究设计。该数据集通过半自动方式从互联网收集，大小为4.1G，适用于深度卷积神经网络的训练。