Compact road-object detection with self-calibrated perspective alignment and quality-aware assignment

Small and distant road objects occupy few pixels, yet preserving fine-resolution features enlarges the candidate field and can destabilize learning under severe class imbalance. We present SC-CADA-YOLO26n, a compact four-scale detector that couples a stride-4 prediction head with bounded self-calibrated perspective-conditioned alignment at P₂ and P₃ , perspective-quality scale-and-rarity-aware target assignment, and mass-normalized foreground reweighting. The two feature-level horizon parameters are learned globally and shared across images. KITTI and BDD100K were each divided into training, validation, and test sets using an 80/10/10 ratio, with validation used for checkpoint selection. At 1280×1280 resolution, the KITTI one-to-many path achieved mAP50:95=0.7534±0.0014 across three training seeds (95% CI [0.7500, 0.7568]). Small-object AP was 0.6720±0.0168 [0.6302, 0.7137], compared with 0.5743±0.0045 [0.5631, 0.5856] for the unmodified baseline. A separate single-run BDD100K experiment achieved mAP50:95=0.325 . The native one-to-one end-to-end path attained 0.7460±0.0023 [0.7402, 0.7517] on KITTI and 0.319 on BDD100K without non-maximum suppression. Intervals describe variation across training seeds on the fixed KITTI test set. Complexity profiling measured 2.547 million parameters and 7.215/28.859 counted GFLOPs at 640×640 / 1280×1280 for the complete unfused one-to-many graph. The results show that coordinating fine-resolution prediction with geometric alignment and quality-constrained supervision yields a compact, accurate detector.