COPING WITH INCREASED LEVELS OF LABEL NOISE IN FACIAL EXPRESSION RECOGNITION

S Balasubramanian, Darshan Gera, Badaveti Naveen Siva Kumar, Bobilli Veerendra Raj Kumar

Abstract: An important problem in facial expression recognition (FER) is the presence of inaccurate annotations referred to as noisy annotations in the datasets. These noisy annotations are present in the datasets due to the subjectivity of annotators, poor quality of images, etc. In our work, we use a sample selection method based on a dynamic adaptive threshold to separate confident samples from non-confident ones. We impose consistency in the negative classes of the non-confident samples to effectively use all the samples. Unlike other methods which consider low label noise (10–30%), we have considered synthetic label noisy datasets with a higher rate of label noise (up to 80%) and demonstrated the proposed framework’s effectiveness using quantitative as well as qualitative results. Our method performs better than the SOTA model at higher rates of label noise in the range of 0.1–3.36% on RAFDB and 0.19–5.84% improvement in FERPlus.