優化亮度感知損失的動態臉部表情識別法

Abstract

none

Facial Expression Recognition (FER) is a meaningful field of research in computer vision and its development could enhance human-computer interactions. Although, some models have been performing very well in laboratory conditions where the intensity of the expression is big and constant, they are not performing that well when put in ‘in-the-wild’ conditions which is closer to real life situations. Dynamic Facial Expression Recognition (DFER) models try to tackle this issue with recognition task on video sequences closer to natural scenes. In those scenes, the intensity of the expression varies a lot and can lead to a bias in the model caused by large intra-class and small inter-class differences. To tackle this issue, the Intensity Aware Loss (IAL) was adopted and it helps the model put extra attention on low intensity samples to prevent the confusion that they may cause. The principle of intensity is illustrated by the difference between the target logit xt and the largest logit excluding the target xmax. When the intensity is low, the expression is more likely to be to confused as another one as they all tend toward the neutral expression when the intensity tends toward 0. Whereas, when the intensity is big, the expressions are easy to differentiate. That is why the intensity can be illustrated by the difference between xt and xmax. A big difference means that the prediction is clear and therefore that the intensity is high, whereas a small difference would mean that the model cannot fully distinguish the expression and therefore that the intensity is low. Using this concept, We thought of using the Euclidian distance between xt and xmax to quantify the intensity of the expression. The idea is that by combining the Euclidian distance with a negative log, the new IAL would be able to emphasize more on the low intensity sample than the original IAL. In this work, we will study the influence of this new IAL while varying its effect on the model over the epochs and combining it with a pretrained model.