Comparison of Manual and AI-assisted Labeling Techniques in Pixel-wise Instance Segmentation

Instance segmentation in computer vision poses significant challenges in dataset preparation, especially for tasks with many instances per image. This is primarily due to the extensive labeling requirements. Recent advancements in deep learning have introduced AI tools, such as the Segment Anything Model (SAM), which can aid in generating accurate segmentation masks. This paper investigates the efficiency and quality of three annotation techniques: manual, interactive AI-assisted, and AI-assisted with pre-labeling, applied to complex instance segmentation of densely packed objects. Our experiments use mushroom segmentation as a representative many-instance problem. Evaluation metrics, including IoU, Precision, Recall, and F1 score, are used to assess annotation accuracy and speed. The results indicate that while manual labeling remains highly accurate for simple tasks, AI-assisted tools significantly improve labeling efficiency and maintain high accuracy for complex tasks with numerous objects. These findings highlight the potential of SAM-based AI-assisted labeling for practical applications in dataset preparation for dense instance segmentation tasks.