Multi-view dynamic facial action unit detection

We propose a novel convolutional neural network approach to address the fine-grained recognition problem of multi-view dynamic facial action unit detection. leverage recent gains in large-scale object by formulating task predicting presence or absence specific still image human face as holistic classification. then explore design space our considering both shared and independent representations for separate units, also different CNN architectures combining color motion information. move setup FERA 2017 Challenge, which we extension that operates first viewpoint from video was taken, evaluating an ensemble detectors were trained viewpoint. Our is holistic, efficient, modular, since new units can be easily included overall system. significantly outperforms baseline with absolute improvement 14% on F1-metric. Additionally, it compares favorably against winner Challenge.