{Home} {Publications and Codes} {Research} {Projects} {Professional Activities} {People}

Machine Vision, Image Recognition and Machine Learning

[Coarse-to-Fine Deep Transfer Adaptation for Kinship]

Paper:

Qingyan Duan, Lei Zhang*, and Wangmeng Zuo, From Face Recognition to Kinship Verification: An Adaptation Approach, IEEE Int'Conf. Computer Vision (ICCV), in AMFG Workshop, 2017. [paper], [Github source code]

Abstract:

Kinship verification in the wild is a challenging yet interesting issue, which aims to determine whether two unconstrained facial images are from the same family or not. Most previous methods for kinship verification can be divided as low-level hand-crafted features based shallow methods and kin data only trained convolutional neural network (CNN) based deep methods. Worthy of affirmation, numerous work in vision get that convolutional features are discriminative, but bigger data dependent. A fact is that for a variety of data-limited vision problems, such as limited Kinship datasets, the ability of CNNs is seriously dropped because of overfitting. To this end, by inheriting the success of deep mining algorithms on face verification (e.g. LFW), in this paper, we propose a Coarse-to-Fine Transfer (CFT) based deep kinship verification framework. As the idea implied, this paper tries to answer “is it possible to transfer a face recognition net to kinship verification?”. Therefore, a supervised coarse pre-training and domain-specific ad hoc fine re-training paradigm is exploited, with which the kinrelation specific features are effectively captured from faces. Extensive experiments on benchmark datasets demonstrate that our proposed CFT adaptation approach is comparable to the state-of-the art methods with a large margin.

Approach:

Illustration of our proposed Coarse-to-Fine Transfer for Deep Adaptation (CFT)

Intuitive illustration of NRML proposed by Lu Jiwen et al. (a) High-dimensional feature space. The data points in blue and red denote parents and children, respectively. The data points with kinship relation are denoted as circles. The data points in the neighborhood and nonneighborhood are denoted as triangles and squares, respectively. (b) The new NRML subspace, where a kinship margin is obtained.

Experiments:

In experiments, two kinds of databases are considered: large-scale CASIA WebFace data (500K) and small-scale KinFace data (4K). The KinFace data include four publicly available datasets, such as KinFaceW-I, KinFaceW-II, Cornell KinFace and UB KinFace.

Convergence curve of coarse-to-fine transfer

References:

[1] J. Lu, X. Zhou, Y.-P. Tan, Y. Shang, and J. Zhou. Neighborhood repulsed metric learning for kinship verification. IEEE Transactions on Pattern Analysis & Machine Intelligence, 36(2):331–345, 2014.

[2] Y. Bengio. Deep learning of representations for unsupervised and transfer learning. ICML Unsupervised and Transfer Learning, 27:17–36, 2012.

[3] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge & Data Engineering, 22(10):1345–1359, 2010.

[4] H. Yan, J. Lu, W. Deng, and X. Zhou. Discriminative multimetric learning for kinship verification. IEEE Transactions on Information forensics and security, 9(7):1169–1178, 2014.