Point Supervised

Point-supervised learning tackles the challenge of training deep learning models for various computer vision tasks using only sparse point annotations instead of expensive pixel-level labels. Current research focuses on developing novel architectures and algorithms, such as those employing self-training with pseudo-label refinement, dense regression of center-directions, and transformer-based approaches, to effectively leverage these limited annotations. This approach significantly reduces annotation burden, making advanced techniques like semantic segmentation, object detection, and action localization accessible for datasets where full annotation is impractical or infeasible, thereby impacting various fields including remote sensing and medical image analysis.