Semantic segmentation of textured mosaics

Abstract This paper investigates deep learning (DL)-based semantic segmentation of textured mosaics. Existing popular datasets for mosaic texture segmentation, designed prior to the DL era, have several limitations: (1) training images are single-textured and thus differ from multi-textured test images; (2) textures typically cut out same raw images, which may hinder model generalization; (3) each image has its own limited set forcing an inefficient one per few data. We propose two datasets, based on existing Outex DTD that suitable networks address above SemSegOutex focuses materials acquired under controlled conditions, SemSegDTD visual attributes in wild. also generate a synthetic version via synthesis can be used way as standard random data augmentation. Finally, we study performance state-of-the-art DeepLabv3+ is excellent variable SemSegDTD. Our allow us analyze results according type material, attributes, various acquisition artifacts, natural versus aspects, yielding new insights into possible usage recent technologies analysis.