Multi-Label Dilated Recurrent Network for Sequential Face Alignment

Tong Yang, Shizheng Qin, Junchi Yan, Wenqiang Zhang

Abstract

Compared with detection in still image, sequential face landmark detection in video is relatively less studied. In this work, we present a novel network with a dilated residual network and a residual convolutional LSTM to preserve the detection acuity at both spatial and temporal dimensions respectively. We also introduce and compare multi-label loss with regression loss and multi-class loss for face landmark detection. We perform extensive experiments to verify the contribution of different components. Our method achieves state-of-the-art performance on benchmark datasets.
BibTeX
@inproceedings{yang2018multi,
  author = {Tong Yang and Shizheng Qin and Junchi Yan and Wenqiang Zhang},
  title = {{Multi-Label Dilated Recurrent Network for Sequential Face Alignment}},
  booktitle = {ICME},
  year = {2018},
  doi = {10.1109/icme.2018.8486438},
}

Loading related content...

Figure