A Semi-Automatic Labeling Approach for Image Segmentation Using YOLOv11

Manual annotation creation for instance segmentation represents one of the main bottlenecks in building datasets for training computer vision models, often being more costly and time-consuming than the model development itself. The proposed method in this work employs an iterative process that aims to improve labeling efficiency, and it consist on the following steps: (i) manual annotation of an initial reduced subset of images, (ii) training of a YOLOv11-seg model, and (iii) automatic inference on unannotated images for generating the remaining masks. Experimental validation was conducted on two public datasets: KITTI Tracking and MADS. A total of 187 models were trained using cumulative training divisions ranging from 0.1% to 100% of available data. The results demonstrate that models can achieve competitive performance even with minimal amounts of data. The analysis reveals that the most significant gains occur when expanding from smaller divisions (1% to 10%), while beyond 50% of data, marginal returns decrease substantially, following the law of diminishing returns. The use of this pipeline proved capable of reducing the required amount of manual annotation creation by up to 20 times. Manual annotation creation for instance segmentation represents one of the main bottlenecks in building datasets for training computer vision models, often being more costly and time-consuming than the model development itself. The proposed method in this work employs an iterative process that aims to improve labeling efficiency, and it consists of the following steps: (i) manual annotation of an initial reduced subset of images, (ii) training of a YOLOv11-seg model, and (iii) automatic inference on unannotated images for generating the remaining masks. Experimental validation was conducted on two public datasets: KITTI Tracking and MADS. A total of 187 models were trained using cumulative training divisions ranging from 0.1% to 100% of available data. The results demonstrate that models can achieve competitive performance (mAP@0.5 > 0.80) even with minimal amounts of data. The analysis reveals that the most significant gains occur when expanding from smaller divisions (1% to 10%), while beyond 50% of data, marginal returns decrease substantially, following the law of diminishing returns.