Deep Features for Recognizing Disguised Faces in the Wild
Abstract : 提出一种基于DCNN的方法,在伪装之后的脸部图像中识别出冒充者。文章在静态图像和视频帧的大数据集。5.6 million images. 上训练两个deep CNNS 不同的网络,该数据集具有 L2-softmax loss 特征。数据集: Wild challenge dataset https://ibug.doc.ic.ac.uk/resources/300-W/
要点:softmax(易实施(现有函数库、数量无限制)收敛很快,但在训练集可能出现过拟合,其适合于high qulitiy data,忽略了mini-batch 的 hard sample) 采用 L2-constrained softmax loss
Face Verification with Disguise Variations via Deep Disguise Recognizer
Abstract : 迁移学习,先利用 inherent face representation 来训练 Residual Inception network framework with center loss
然后迁移到下一个网络,利用 Inception-ResNet model using disguised faces 来训练,以减少disguise 对识别的影响。
尽可能近地嵌入同一特征,由此可将个体的disguise 进行聚类。
Inception-Net based features for verifying faces with disguise variations
表达:In this paper, we propose a novel deep learning based framework for enhancing the performance of automatic face recognition with disguised faces.
Disguised Faces in the Wild
Abstract : This paper presents a description of the dataset, the baseline protocols and performance, along with the phase-I results of the competition.
对这一数据集的介绍,并且包括目前在这个比赛(领域)上的阶段性成果