CVAT meets transformers: accelerating semantic segmentation labeling in industrial applications

Semantic segmentation plays a crucial role in various industrial applications by enabling detailed analysis and understanding of images at the pixel level. This thesis presents a novel approach to streamline the semantic segmentation labeling process by integrating Computer Vision Annotation Tool (CVAT) with advanced Transformer models, specifically focusing on industrial applications. Traditional methods for preparing datasets for semantic segmentations relying on manual annotation are time-consuming and economically burdensome. To address these challenges, we explore the integration of CVAT, an online tool designed for efficient image labeling, with Transformer-based models. Our methodology involves a semi-automatic pipeline leveraging Segment Anything Model (SAM) within CVAT for initial annotations, followed by fine-tuning of YOLO , SegFormer, and the ViT-Adapter which then can be used to label new images to help the labeler in his/her work. We detail the process of fine-tuning these models and we discuss about the improvements in terms of time saving.