Research on data construction for tank target detection

In view of the current problem in tank target detection research, where the focus is on algorithm optimization while neglecting the practical effectiveness of data, this study for the first time systematically examines the application efficacy of two types of data in real-combat scenarios. One type consists of publicly available datasets and internet-sourced images, while the other comprises tank images constructed through expert-curated selection. Through the creation of five distinctive datasets and conducting comparative experiments using the YOLOv10 algorithm, this study focuses on assessing its detection performance in real-world combat conditions. Findings reveal that the model trained on publicly sourced data exhibits extremely weak generalizability in real-combat scenarios, achieving a highest detection accuracy of only 13.60%, highlighting the limitations of generic data when applied to specialized contexts. In contrast, the dataset constructed based on expert knowledge shows significant advantages. After optimizing the model complexity, it achieves a real-combat detection accuracy of 41.70%, marking a remarkable improvement. This study quantifies the performance gap between publicly sourced data and expert-curated data, underscores the critical role of expert knowledge in data construction, and provides valuable insights for data engineering practices in military target detection.