a51fb975227d6640e4fe47854476d133-AuthorFeedback.pdf

The comparison to LP-SM is consistent. Facial expression can be a group of AUs. While AU annotations have less31 personal bias, they are noisy because the motion is subtle and difficult to distinguish even for experts. Expression32 is global and easier to label but it is subject to personal bias. The learning of rBN requires no training data, i.e., no expression and AU labels from a37 specificdataset.