Improving Performance of Instance Segmentation Model for Building Object Detection Using Contrastive Unpaired Translation

With the advancements in data collection processes and sensors, a vast amount of data is now available, driving the increasing application and utilization of deep-learning-based artificial intelligence technologies. For instance, object detection through image data is utilized in various fields such as traffic safety, crime prevention and public safety, environmental monitoring, and disaster response in urban areas. Deep-learning-based object detection models exhibit high performance, but there are limitations in performance when the training data is restricted. There are issues with degraded object detection performance owing to differences in environmental factors (occlusion and illumination) in the data collected under different solar altitudes or shooting conditions from the training data. The aim of this study is to enhance the performance of object segmentation by mitigating different environmental factors between the training and collected data using the contrastive-unpaired-translation (CUT) algorithm, one of the image-to-image translation algorithms. In this study, we aim to improve object segmentation performance by generating images under environmental conditions (e.g., shadows and shading) similar to those of the training data. The object segmentation model used in this study was You Only Look Once version 8, and instance segmentation was performed by inputting data with and without applying CUT. The results showed an improvement of approximately 11.11% in mAP@50. Furthermore, statistical verification confirmed that this is a significant difference. The results of this study confirmed the potential of improving instance segmentation performance through image translation techniques, which will contribute to autonomous driving and unmanned services