search

nxsEdson / MLCR

Multi-label Co-regularization for Semi-supervised Facial Action Unit Recognition (NeurIPS 2019)
63 stars 18 forks source link

Question About mv-loss #8

Closed ky941122 closed 3 years ago

ky941122 commented 3 years ago

Hi, I have recently read your MLCR paper and am very interested in your ideas. But I'm a little confused about the mv-loss. Why don't you just constrain the features of different views to be mutually orthogonal, but go for the classifier's weights of different views instead?

nxsEdson commented 3 years ago

Sorry for the late reply, and thank you for your interests in our work. There are two reasons for using the weights instead of the features: 1) the features are shared and used for multilabel classification, e.g., 12 AU classification tasks, and 2) the weights for each AU can be regarded as a representation of the features for this specific AU because the weights vector can be seen as the base vector for projection, and the projection results are the final prediction.