Continual Online Learning For Unconstrained Facial Landmark Detection And Tracking

Facial landmark localization is of profound interest in computer vision since it is critical in solving many key affective computing problems such as face recognition, expression recognition, emotion detection etc. Face landmark alignment, which is all about localization of key facial structures such as eyes, nose, mouth etc., is quite challenging in nature owing to wide range of face appearance variations. This is mainly due to various head poses, environmental lighting conditions, external occlusions, and imaging sensor noise patterns. Most important challenge in building the landmark detection and tracking solution stems from the fact that it is difficult to train a machine learning model with training data that fully spans all the aforementioned attribute space. Hence by enabling the model to learn incrementally from test time environment, this challenge can be addressed to a significant extent. However, incremental online learning often results in a phenomenon known as catastrophic interference i.e. when the model learns new information it starts forgetting the previously learned information. Continual learning approaches are demonstrated to be capable of handling this interference effectively in case of connectionist learning procedures. In this proposal the possibility of employing the continual learning methods to address the catastrophic forgetting phenomenon during incremental training of landmark detection and tracking models is discussed.