Face alignment by learning from small real datasets and large synthetic datasets
Face alignment by learning from small real datasets and large synthetic datasets
复制标题
DOI：
10.1109/cacml55074.2022.00073
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
Haoqi Gao;K. Ogawara
中科院分区：
文献类型：
--
作者：
Haoqi Gao;K. Ogawara
近年来，与面部相关的研究在现实生活中得到了广泛的应用。但其应用引发的隐私侵犯、数据滥用等问题也引发全球争议。不可否认，人脸技术高效便捷，但人脸技术隐藏的危险和风险也应综合考虑。当前的人脸算法在复杂且具有挑战性的环境（例如大角度或表情）中仍然具有挑战性。首先，现有的公共训练数据集大多是正面，具有挑战性的数据分布不平衡。其次，收集的真实数据集需要明确的用户同意，并且注释过程耗时且昂贵。在本文中，我们通过合成数据集开辟了一个新的研究方向。我们尝试使用合成数据集来减少模型对现实世界数据集的依赖。人脸对齐实验探索了合成数据集的互补性和可用性。
In recent years, face-related research had a wide variety of real-life applications. However, issues such as privacy violations and data abuse caused by its applications have also triggered global controversy. It is undeniable that face-related technology is efficient and convenient, but the dangers and risks are hidden by face technology should be comprehensively considered. Current face algorithms are still challenging in complex and challenging environments (e.g., large angles or expressions). Firstly the existing public training datasets are mostly frontal faces, which have an unbalanced distribution of challenging data. Secondly, the collected real datasets require explicit user consent, and the annotation process is time-consuming and expensive. In this paper, we open a new research direction through synthetic datasets. We try to use synthetic datasets to reduce the dependence of the model on the real-world data set. The face alignment experiments explore the synthetic dataset's complementarity and availability.