Hidden Variables in Real-Time Person Detection: Loss-Function Screening, Annotation Geometry, and Operating-Point Selection in YOLOX-Based Single-Class Detectors

Person detection underpins a growing range of commercial systems, from retail analytics to safety monitoring, yet the experimental practices used to configure such detectors are rarely examined. This paper presents an end-to-end empirical study of single-class person detection built on YOLOX and trained on the person subset of COCO 2017 (64,115 images; 262,465 instances), stress-testing three routine practices: (i) short-run screening of bounding-box regression losses, (ii) evaluation under a fixed confidence threshold, and (iii) implicit reliance on a single annotation geometry. All three are found to be unreliable in ways that are easy to miss with a single-number comparison. Nine regression losses screened for 15 epochs on a 1000-image subset differ by only 1.06 AP, and their ranking shows no significant association with an independent benchmark (Spearman’s ρ=−0.23, p=0.55, n=9), providing no evidence that short-run screening predicts external performance. Under full 100–120-epoch schedules, the DIoU configuration reaches +1.4 AP over IoU within a fixed annotation regime—under a 20-epoch-longer schedule; so, the margin is not attributable to the loss alone—whereas switching the training and evaluation boxes from original COCO to segmentation-derived tight boxes shifts AP by up to 3 points, confounding naive cross-run comparisons; a separate SIoU+focal configuration diverges. Threshold sweeps against an off-the-shelf YOLO11m show that the better detector flips with the operating point: YOLO11m peaks at F1 = 0.870 at low thresholds, while the fine-tuned YOLOX-M retains F1 = 0.827 versus 0.751 at the common τ=0.5. Finally, hardware benchmarking (0.9 M–33.7 M parameters, 2.5–252 FPS measured directly on our hardware) and a containerized RTSP pipeline translate the findings into deployment guidance. Reporting operating-point curves, annotation geometry, and screening protocols, we argue, is as important as the loss function itself.