Coarse Annotation

Coarse annotation, a technique using less precise labels than pixel-perfect annotations (e.g., bounding boxes instead of pixel-level masks, or single points instead of keypoints), is gaining traction in various computer vision tasks. Current research focuses on developing methods to effectively leverage these coarser labels for training accurate models, often employing techniques like semi-supervised learning, multiple instance learning, and self-training frameworks to refine initial annotations or generate pseudo-labels. This approach significantly reduces annotation costs, making large-scale data collection and model training more feasible for applications like medical image analysis, urban scene understanding, and object detection, ultimately improving the efficiency and accessibility of deep learning in diverse fields.