Transferring Landmark Annotations for Cross-Dataset Face Alignment

 

 

Shizhan Zhu^1*, Cheng Li^2*, Chen Change Loy¹, Xiaoou Tang¹

 

¹ Department of Information Engineering, the Chinese University of Hong Kong

² Department of Physics, Tsinghua University

* denotes equal contributions

 

Abstract

 

Dataset bias is a well-known problem in object recognition domain. This issue, nonetheless, is rarely explored in face alignment research. In this study, we show that dataset plays an integral part of face alignment performance. Specifically, owing to face alignment dataset bias, training on one database and testing on another or unseen domain would lead to poor performance. Creating an unbiased dataset through combining various existing databases, however, is non-trivial as one has to exhaustively re-label the landmarks for standardisation. In this work, we propose a simple and yet effective method to bridge the disparate annotation spaces between databases, making datasets fusion possible. We show extensive results on combining various popular databases (LFW, AFLW, LFPW, HELEN) for improved cross-dataset and unseen data alignment, which is challenging without the proposed method.

 

 

 

Demo:

 

 

Downloads:

We use the annotation transferring approach to provide dense landmarks annotation (68 and 194 pts per face) on the famous verification dataset LFW based on its original 5-point annotation.

 

We manually check for the accuracy of the landmarks annotation.

Overall performance:

 

   Download 68pts and 194pts dense landmarks annotation with detailed manually checking results for LFW (dense-LFW) from here.

 

References:

Shizhan Zhu, Cheng Li, Chen Change Loy, Xiaoou Tang. Transferring Landmark Annotations for Cross-Dataset Face Alignment, Technical Report, arXiv:1409.0602, 2014.