Abstract

Driven by the advancement in technology that can facilitate implementation of deep neural networks (DNNs), and due to the availability of large scale datasets, automatic recognition performance of the machines has increased by leaps and bounds. This is also true with regard to facial expression recognition (FER) wherein the machine automatically classifies a given facial image in to one of the basic expressions. However, annotations of large scale datasets in FER suffer from noise due to various factors like crowd sourcing, automatic labelling based on key word search etc. Such noisy annotations impede the performance of FER due to the memorization ability of DNNs. To address it, this paper proposes a learning algorithm called Co-curing: peer training of two joint networks using a supervision loss and a mimicry loss that are balanced dynamically, and supplemented with a relabeling module to correct the noisy annotations. Specifically, peer networks are trained independently using supervision loss during early part of the training. As training progresses, mimicry loss is given higher weightage to bring consensus between the two networks. Our Cocuring does not need to know the noise rate. Samples with wrong annotations are relabeled based on the agreement of peer networks. Experiments on synthetic as well real world noisy datasets validate the effectiveness of our method. State-of-the-art (SOTA) results on benchmark in-the-wild FER datasets like RAF-DB (89.70%), FERPlus (89.6%) and AffectNet (61.7%) are reported.