Reducing Manual image annotation effort using sam, dinov3 and active learning

Sammanfattning

Modern computer vision systems rely heavily on annotated image datasets, but creating object-level annotations is both time-consuming and expensive. This thesis investigates how active learning can be integrated into a semi-automatic annotation workflow in order to reduce the amount of manual labeling while still achieving reliable classification performance. The proposed workflow combines the Segment Anything Model (SAM), DINOv3-based feature extraction, active learning, and a graphical interface within a unified annotation pipeline. SAM is used to extract object regions, while DINOv3 generates feature representations for classification. Active learning is then applied to iteratively select informative samples using both random sampling and uncertainty-based sampling strategies. The experiments were conducted on a subset of the Microsoft COCO dataset using four experimental configurations: SAM with random sampling, SAM with uncertainty sampling, DINO with random sampling, and DINO with uncertainty sampling. Fully automatic segmentation was also compared with box-prompted segmentation to evaluate how segmentation quality affects the overall workflow. The experiments showed that classification performance improved as additional labeled samples were introduced. The guided segmentation setting achieved a mean IoU of 0.879 together with precision, recall, detection F1-score, and AP@0.5 values of 1.000. In the active learning experiments, DINO with uncertainty-based sampling achieved stronger performance during the early annotation stages, while SAM with random sampling achieved the highest final macro F1-score of 0.705 at larger annotation budgets. The experiments also showed that fully automatic segmentation generated several irrelevant object proposals, whereas box-prompted segmentation improved localization reliability and annotation quality. The findings suggest that active learning can improve annotation efficiency, but that the overall performance depends strongly on segmentation quality and feature representations.