使用分类CNN进行人脸属性预测.pdfVIP

Face Attribute Prediction with Classification CNN Yang Zhong, Josephine Sullivan, Haibo Li Computer Science and Communication KTH Royal Institute of Technology 100 44 Stockholm, Sweden Comments: In proceedings of 2016 International Conference on Image Processing (ICIP) Subjects: Computer Vision and Pattern Recognition Outline • The key to attribute prediction is to build proper feature representations to cope with these unfavorable conditions • The face attributes are different between local attributes and global attributes • Use the mid-level CNN features instead of high-level ones for attribute prediction • The mid-level representations are robust to face recognition and attribute prediction Discussion on CNN representations • The feature from high-level FC layers may loss some spatial information • Different levels of Con encode different levels of ion • The earlier CNN layers are likely to better describing local facial attributes • The later CNN layers have a high chance to learn global facial attributes The Major Contributions • Reveal the fact of the diverse utilities of deep hierarchical representations for face attributes • Show the effective performance on predicting describable face attributes • Improve the state-of-the-art on two large scale datasets with the mid-level representations • Only constru ingle dee work for face recognition and attribute prediction The Structure of CNN • All the resulting representations from convolutional layers have same feature map size of • Prediction effectiveness discrepancy of hierarchical representations over 40 attribu