| Title |
Quantitative Error Detection and Categorization for SAM2-based Multiple Object Tracking |
| Authors |
최주은(JuEun Choi) ; 이주현(JuHyeon Lee) ; 이승현(SeungHyun Lee) ; 정경훈(KyeongHoon Jung) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.2.103 |
| Keywords |
Multiple object tracking; Image segmentation; SAM2; Overlap error; Exchange error |
| Abstract |
With the growing demand for video-based recognition systems, the importance of Multiple Object Tracking (MOT) technology has become increasingly important. Recently, image segmentation techniques have been integrated into MOT to achieve more precise object separation, with general-purpose segmentation models such as Segment Anything Model 2 (SAM2; Segment Anything in Images and Videos) being widely utilized. Although SAM2 performs well across diverse application domains, its segmentation accuracy deteriorates in scenes where similar objects overlap or blend into complex backgrounds. In this study, we propose a method to automatically detect segmentation error events in frames generated by SAM2-based object segmentation, thereby addressing these limitations. The proposed method categorizes error types in overlapping-object scenarios into duplication errors and exchange errors. It quantitatively detects them using object mask area variations, overlap ratios, and bounding box intersections. Validation on eight experimental videos with various resolutions and environments demonstrated that the proposed algorithm detected segmentation error and confirmed error correction through prompt re-input. Overall, we provide a practical and effective solution to enhance the reliability and scalability of SAM2-based MOT systems in complex environments with overlapping objects. |