Computer Vision News - November 2021

65 ... on semi-supervised learning at UCL augmentations. By using the standard deviation of the mean of all outputs for one image, the certainty can be derived. Dataset The dataset I used is publicly accessible here: https://weiss-develop.cs.ucl.ac.uk/ and is provided by the WEISS center itself. It consists of manually annotated frames and frames with predictions obtained with leave-one-out cross-validation and was combined with additional in-house unlabelled images. I used the annotated images for the initial training and then generated pseudo masks for the unannotated frames to enhance the dataset for further training the self- teaching model and for teaching the student model. Example fetoscopy image with manually annotated ground truth mask. My implementation I implemented two simple semi-supervised versions, a self-teaching network, and the second one followed the student-teacher approach. I used the U-Net architecture with a ResNet-50 backbone for PyTorch which is available on GitHub.