Semantic segmentation is the process which aims to classify individual pixels of an image. Recently, Kaggle hosted the 2018 Data Science Bowl competition dedicated to nucleus detection and segmentation based on microscopic images. In this talk, I will present two approaches to this problem, based on U-Net and Mask R-CNN.
U-Net is a deep learning approach to image segmentation task, that works with a relatively small number of training images and produces precise segmentation. On the other hand, Mask R-CNN requires much more training data, however, it detects many object instances and simultaneously generates high-quality segmentation map for each object. While Mask R-CNN outperforms all existing, single-model entries on every task (including the COCO 2016 challenge winners), it has lost 2018 Data Science Bowl competition with much simpler U-Net with heavy pre and postprocessing. In this talk, I would like to concentrate on the reasons for this failure. Moreover, I will present the possible extensions of both methods and show the practical guidelines of how to use them.