Глазкова Екатерина Васильевна
Semantic Image Segmentation With Deep Structured Models
Прикладная математика и информатика
Recent deep learning models for semantic segmentation contain convolutional parts with pixel-wise predictions and mainly do not directly employ interactions between pixel labelling results. Structured models eliminate this disadvantage and explicitly use interactions between image pixels. We use Conditional Random Field (CRF) structured model as a separate model and as a part of an end-to-end trainable model. We investigate how it might be implemented as a deep network layer and experiment with Convolutional CRF and Deep Gaussian CRF deep structured models. We use UNet as a base Convolutional Neural Network (CNN) model. We consider joint and separate training of deep structured models, structure modifications and implementation simplifications. We show that structured models are superior to only pixel-wise predictions and compare the considered models by time and quality of predictions.