Fully-Automatic Facial Expression Recognition (FER) from still images is a task as it involves handling large interpersonal morphological, and as partial occlusions can occasionally happen. Furthermore, expressions is a time-consuming process that is prone to, thus the variability may not be fully covered by the training. In this work, we propose to train Random Forests upon spatially defined subspaces of the face. The output local predictions form a categorical-driven high-level representation that we call Local Expression (LEPs). LEPs can be combined to describe categorical facial as well as Action Units (AUs). Furthermore, LEPs can be weighted by scores provided by an autoencoder network. Such network is trained locally capture the manifold of the non-occluded training data in a way. Extensive experiments show that the proposed LEP yields high descriptive power for categorical expressions and AU prediction, and leads to interesting perspectives towards the design occlusion-robust and confidence-aware FER systems.
No takes yet. Share an insight, caveat, or question.
Dapogny et al. (2016) studied this question.