Web-Scale Training for Face Identification

Web-Scale Training for Face Identification CVPR2015

Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato
Facebook AI Research
Menlo Park, CA 94025, USA
fyaniv, mingyang, ranzatog@fb.com
Lior Wolf
Tel Aviv University
Tel Aviv, Israel
wolf@cs.tau.ac.il

Scaling machine learning methods to very large datasets.
We study face recognition and show that three distinct properties have surprising effects on the transferability of deep convolutional networks (CNN):
(1) The bottleneck of the network serves as an important transfer learning regularizer.
bottleneck层是重要的迁移学习正则层
(2) in contrast to the common wisdom, performance saturation(饱和) may exist in CNN’s (as the number of training samples grows) ; we propose a solution for alleviating(缓解) this by replacing the naive random subsampling of the training set with a bootstrapping(引导) process.

bootstrapping procedures aim at focusing on the hardest cases ignoring or down-weighing the easy ones.

用一个引导过程代替对训练集作的简单随机下采样，可以缓解由训练集增大带来的训练过程饱和
(3) we find a link between the representation norm and the ability to discriminate in a target domain, which sheds lights on(一些启发) how such networks represent faces.

找到表达准则和特定领域判别能力的关联，给表示脸带来一些启发
We discover a three-way link between the representation norm, the image quality, and the classification confidence.

Face verification, is the task of determining whether two face images belong to the same subject.

训练集10 million subjects with 50 images each in average

数据集Bootstrapping方法：
从10M人里随机挑选100人作为种子，再为每个种子挑选1000个最接近的人。两个人之间的相似度以角度的cosine值为超参数

所有的的55000个人的图片构成新的数据集DB2。
数据集特点：包含简单和复杂的样本，不同种子之间的辨识很容易，但是每个种子的邻近辨识很困难。

实验

两个方法：
1.在封闭的验证集上，gallery set有4249个人，每人一张图片。probe set有3143张图片，都是gallery set里面的人。
在这里用Rank-1 identification accuracy来表现结果。
2.在开放的数据及上，gallery set有596个人，每人一张图片。probe set有596个真人（对应gallery set）和9491个假货。
在这里用Rank-1 Detection and Identification Rate (DIR)来表现结果。
which is the fraction of genuine probes matched correctly in Rank-1 at a 1% False Alarm Rate (FAR) of
impostor probes that are not rejected.应该是Rank-1正确配对的除以当FAR在1%时的错误图片的比率。

一、压缩实验

二、引导表示Bootstrapped Representations

比较在55K人，4.5M图片（bootstrapped identities）上学到的特征和随机挑选的108K人，3.2M图片以及250K人，7.5M图片。

结论：
1.DNN训练数据越多效果越好；
2.引导过的55K数据集训练效果比较好；
3.扩大局部连接层后泛化性更好，但是更大的网络效果不一定好。
An even larger improvement is
obtained when the locally-connected layers (L4-L5-L6) are
expanded as described in Sec. 3.4, and the extended 256D
and 1024D representations (denoted as 256+ and 1024+)
generalize better than their unmodified counterparts.
三、与现有成果比较

四、模型选择

利用一个separate internal validation dataset，包含了10000个人。Gallery set有55000图片（平均每人5.5张）。
Probe set 有100000图片，每人有10张queries。
分别测试4096维特征，256D维特征和1024+维特征的cosine similarity。
用Rank-1 accuracy and the DIR at 1% and 0.1% FAR表现效果。
结论：1）压缩过的256D特征泛化性比原始4096特征更好；
2）引导过的1024+特征改善了DIR