2018/10/13 by Pierluigi Zama Ramirez, Alessio Tonioni, Ramirez, Pierluigi Zama +3
Computer Science · Medicine · #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1810.05852
openalex publication_date 2018/10/13 · openalex created_date 2022/08/02 · openalex updated_date 2026/07/28
Performance achievable by modern deep learning approaches are directly\nrelated to the amount of data used at training time. Unfortunately, the\nannotation process is notoriously tedious and expensive, especially for\npixel-wise tasks like semantic segmentation. Recent works have proposed to rely\non synthetically generated imagery to ease the training set creation. However,\nmodels trained on these kind of data usually under-perform on real images due\nto the well known issue of domain shift. We address this problem by learning a\ndomain-to-domain image translation GAN to shrink the gap between real and\nsynthetic images. Peculiarly to our method, we introduce semantic constraints\ninto the generation process to both avoid artifacts and guide the synthesis. To\nprove the effectiveness of our proposal, we show how a semantic segmentation\nCNN trained on images from the synthetic GTA dataset adapted by our method can\nimprove performance by more than 16% mIoU with respect to the same model\ntrained on synthetic images.\n