Computer Vision News - November 2021

64 My summer internship ... • Graph-based (e.g. label propagation) Label propagation is a transductive algorithm that finds groups in a graph and therefore is able to label unlabelled data based on labelleddata. The assumption followed is that data points close to each other belong to the same group. If you are interested in learning more, I’d recommend the paper “ An Overview of Deep Semi-Supervised Learning ” which is available on arXiv . Additionally, both Papers with Code: Semi-Supervised Semantic Segmentation and Semi-supervised learning for medical image segmentation: SSL4MIS on GitHub show recent advances in the semi-supervised world. How to deal with certainty? For the methods using pseudo-labelling or consistency regularisation, certainty is taken into account to filter or weight labels. Certainty is information that can easily be calculated however it is often ignored. To get the model certainty for a sample (not to be confused with the probability output), approaches like Monte Carlo Dropout (dropout at test-time) or test-time augmentation (TTA) can be used. In both cases, an image is put through a network multiple times and leads to different outputs due to the randomly dropped nodes or different Semi-supervised learning using GAN. supposed to generate samples that look similar to the real data while the discriminator needs to learn to distinguish between real and fake samples. This is done to learn features, later on the discriminator is trained in a supervised manner to predict classes of the real examples. Alternatively, the discriminator can be trained simultaneously to predict K+1 classes, including K real classes and the fake class.