Abstract
<title>Abstract</title> <p>Standard object detection evaluation relies on hard-threshold metrics (mAP, F1) at a fixed IoU threshold. Their binary nature produces unstable rankings, penalizes near-correct detections as harshly as complete misses, and ignores semantic label relationships. We propose CCS, a Continuous Spatial-Semantic Concordance Score with two complementary terms: (i) C sp models bounding boxes as 2D Gaussians and computes closed-form normalized cosine similarity , smoothly penalizing spatial deviation without a hard threshold; (ii) C sem uses taxonomy-driven Wu-Palmer similarity to award partial credit when a predicted label is semantically close to ground truth (e.g., “red-spotted tongue” vs. “red tongue”). We benchmark CCS against IoU, GIoU, DIoU, NWD, and COCO mAP@[0.5:0.95] across six YOLO variants (v8/v10/v11) on an eight-class tongue-symptom dataset with severe per-class difficulty imbalance. F1@0.5 ranges 0.47–0.50 and degrades sharply at stricter thresholds; CCS stays stable at 0.62–0.65. C sem (0.67–0.70) exceeds C sp (0.57–0.60), showing semantic proximity rescues clinically meaningful near-misses. Five model pairs reverse ranking across IoU thresholds, which CCS avoids by construction; Spearman ρ = 0.94 with F1@0.5 confirms CCS preserves overall ranking. CCS ranking is stable for α ∈ [0.5, 0.7]; entropy weighting (α = 0.64) reproduces the default ranking. These results position CCS as a threshold-independent candidate for detection evaluation in class-imbalanced, semantically structured label spaces.</p>