Structured Instance Understanding with Boundary Box Relationships in Object Detection System
Supasate Vorathammathorn, Thanatwit Angsarawanee, Sakol Tasanangam, Theerat Sakdejayont · 2024
The paper introduces Structured Instance Understanding (SIU) with boundary box relationships model, which improves object detection systems by validating the structural relationships of resultant bounding boxes. SIU addresses the critical issue in object detection systems: the misidentification of object parts and the challenges of accurately recognizing structural relationships within bounding boxes. This problem is especially prevalent in complex scenes where objects can be similar, partially obscured, or closely positioned, leading to inaccuracies in detection and classification. SIU operates independently of object detection models as a postprocessing model, and SIU involves two frameworks in model training: 1) synthesis of error-reflective data and 2) feature creation for structural comprehension. The experimental result shows that SIU increases Mean Average Precision (mAP) and F1-scores, effectively distinguishing correct from incorrect detection results, compared to a baseline method where predictions are validated via confidence score thresholding. This confirms SIU's potential to advance image understanding applications, showing significant strides over the traditional confidence score method.