Illu-NASNet: unsupervised illumination estimation based on dense spatio-temporal smoothness

Illumination estimation is a highly challenging problem. Many methods are learning from multi-image by unsupervised learning. However, these methods are based on unrelated data and do not exploit the information between images. In this paper, we propose a new unsupervised method for learning illumination by observing indoor video sequences under changing illumination and learning the intrinsic decomposition of the sequences. This method allows us to learn without ground truth (GT) and leverage information obtained from multiple consecutive images. Based on the above ideas, we propose a new network framework and introduce albedo smoothness loss and illumination smoothness loss two new dense spatio-temporal smoothness loss functions. These loss functions take full advantage of the information between images to constrain the entire image sequence. In our evaluation, our approach shows good performance on several specific metrics. Experiments show that our method has strong generalization and can be easily applied to other classical datasets, including Intrinsic Images in the Wild (IIW) and Shading Annotations in the Wild (SAW).