ECCV 2020 Daily - Thursday

Learning to segment microscopy images with lazy labels by Ke et al. explores the question of how much can be learnt from "lazy" labels, i.e., rough strokes to detect and separate cells instead of precise ground truth segmentation. Using a multi-task learning objective tailored to those labels, the authors show that the use of additional lazy labels greatly reduces the need for precise labels: their method achieves at least the accuracy of a conventional supervised method while using only around 10% of the available precise labels . 3 BioImage Computing 21 The lack of ground-truth can also be compensated for using learnt priors from the images alone, as demonstrated by Buchholz et al. DenoiSeg: Joint Denoising and Segmentation is a method to jointly train a network on a self-supervised denoising task and a supervised segmentation task. Extensive experiments demonstrate that this co-learning allows the network to make better use of the available ground-truth. DAILY T h u r s d a y