Manual data annotation involves human annotators labeling and reviewing data according to predefined criteria. It can be time-consuming and expensive, so automated data annotation based on artificial intelligence algo...
详细信息
ISBN:
(纸本)9783031809453;9783031809460
Manual data annotation involves human annotators labeling and reviewing data according to predefined criteria. It can be time-consuming and expensive, so automated data annotation based on artificial intelligence algorithms has gained popularity. However, the classical integrated artificial intelligence models have some limitations, especially in complex scenarios characterized by occlusions, low resolution, and illumination variability. This paper introduces new evaluation methods for imageinstancesegmentation, focusing on the need for easy-to-understand quality metrics in the face of the complexity of image annotation. Considering the difficulties of manual annotation processes and the diversity of computer vision models focused on instancesegmentation, four quality metrics are proposed: Number of detections, Normalized Bounding Box Area, Normalized Mask Area, and Bounding Box Occlusion Rate. These metrics attempt to overcome the difficulties of select the best model for automated labeling based on conventional metrics such as Intersection Over Union and Mean Average Precision, especially in scenarios where classical models have limitations. The application and analysis of these metrics in the OVIS database demonstrates their potential to improve the interpretation of instancesegmentation models, thus facilitating a more accurate and accessible automated annotation process.
暂无评论