Exploring Loss Function Schemes for Single Image Reflection Removal
Final project for the course CSC2547 - Topics in Machine Learning: Machine Learning for Machine Vision as Inverse Graphics.
Abstract
Loss computation plays a very important role in guiding the training of reflection removal model. Previous studies showed that a well defined loss function can help better remove reflections and generate clearer result images. In this study, we apply several loss functions that are not used in reflection removal domain before. Furthermore, instead of empirically selecting the weights for each loss components, we apply the uncertainty-based auto loss weighing method to dynamically learn the weights of each loss components from the total loss. The experiment shows that by combining the auto loss weighing method with our newly proposed loss function, the SOTA baseline model we use have a significant performance improvement over the synthetic reflection removal dataset.