Abstract

Facial micro-expressions are quick, involuntary and low intensity facial movements. An interest in detecting and recognizing micro-expressions arises from the fact that they are able to show person’s genuine hidden emotions. The small and rapid facial muscle movements are often too difficult for a human to not only spot the occurring micro-expression but also be able to recognize the emotion correctly. Recently, a focus on developing better micro-expression recognition methods has been on models and architectures. However, we take a step back and go to the root of task, the data.

We thoroughly analyze the input data and notice that some of the data is noisy and possibly mislabelled. The authors of the micro-expression datasets have also acknowledged the possible problems in data labelling. Despite this, no attempts have been made to design models that take into account the potential mislabelled data in micro-expression recognition, to our best knowledge. In this paper, we explore new methods taking noisy labels into special account in an attempt to solve the problem. We propose a simple, yet efficient label refurbishing method and a data cleaning method for handling noisy labels. The data cleaning method achieves state-of-the-art results in the MEGC2019 composite dataset.