A semi-automatic annotation method for instance segmentation datasets of industrial objects
Shuai Yang, Xuemei Liu, Yuanxin Gao · Procedia CIRP · 2025
In the industrial sector, it is challenging and expensive to acquire real datasets for object detection and instance segmentation. Furthermore, the detection accuracy of models trained with virtual dataset is short to meet industrial requirement. Neural network detection algorithm is difficult to apply to the industrial sector. To address the problems, a semi-automatic annotation method for instance segmentation dataset of industrial objects was proposed. Firstly, a virtual environment was created based on Unity 3D, which allowed for the automatic generation of the virtual dataset of industrial objects. Secondly, the virtual dataset was utilized to train a neural network model as an annotation tool to automatically annotate the real data. The annotation results were subsequently manually screened to compile the required dataset. Comparison experiments were carried out between the proposed method and manual annotation datasets. The results show that the average time required for single image annotation was 5.2s using the proposed method, which was approximately 9 times improvement compared to 46.1s for manual annotation. Additionally, the mean average precision (mAP) of the model trained with the mixed dataset generated by the proposed method reached 73.61%, which was close to the 76.49% achieved by the model trained with the manual annotation dataset, validating the effectiveness and superiority of the method.