Ambiguous Boundary Uncertainty Reduction in Single Stage Detector Models

The MS-COCO dataset provides ground truth bounding boxes with high accuracy, leaving minimal room for ambiguity. However, during the annotation process, uncertainties can still arise, particularly in situations where object boundaries are indistinct or when numerous annotators are involved. In this paper, we introduce a novel bounding box regression loss that improves the bounding box localization accuracy during the inference phase, particularly in scenarios characterized by the aforementioned uncertainties. During training, we also experiment with a thresholding procedure at the time of loss computation, which leverages the Intersection over Union (IoU) value, consequently enhancing the stabilization and precision of the bounding box localization. We evaluate the novel loss on multiple datasets focusing especially on metal surface defect datasets, yielding favorable results. When tested on the MSCOCO dataset for small object detection, we observe a gain in the Average Recall for small objects ($\mathbf{A R}^{\mathrm{S}}$) and Average Precision with 0.5 IoU value ($\mathrm{AP}^{0.5}$) of YOLOv7 by 1.83% and 1.61% respectively. We also validate the algorithm using other metal defects datasets, namely GC10-DET, and proprietary custom aircraft surface metal defects dataset, with consistently positive outcomes with up to 19.41% gain in Precision, 60.20% gain in Recall, 43.48% gain in mAP@. 5 and 46.86% gain in [email protected]:.95 for some classes.