Architecture, challenges and solutions:
Image segmentation is a classic computer vision problem. We adapted a pretrained convolutional neural network by fine-tuning it on a manually annotated dataset. The results our model got were close to those of human annotators, with our model being able to segment tens of images per second.
