術具位置推定のための画像生成手法に関する研究

抄録

Using nasal endoscopic surgery as a case study, we are investigating the pose estimation of surgical tools using endoscopic images. Current state-of-the-art algorithms are well known to require a large amount of training data. Our group has been investigating the generation of synthetic images with realistic appearance to consistently address this difficulty using photorealistic rendering. In this work, we preliminarily investigate the use of generative-adversarial networks to further improve the rendered images' appearance. Our current implementation reveals some difficulties in this strategy and highlights directions for future work.