<![CDATA[This study benchmarks YOLO26 against YOLOv11 for detecting minority waste categories (Hazardous/B3 and Residue), evaluating whether YOLO26's Progressive Loss Balancing (ProgLoss) and Small-Target-Aware Label Assignment (STAL) mechanisms address class imbalance and small-object detection challenges. Both models were trained under identical conditions (50 epochs, 640×640) on a combined TACO, RecyBat24, and Food Waste Detection dataset (2,326 images, 70:15:15 split), across three scenarios: YOLOv11 baseline, YOLO26 default, and YOLO26 with class weighting and copy-paste augmentation. YOLO26 (default) achieved a marginally higher mAP@0.5:0.95 (0.393 vs 0.385) and modestly faster CPU inference (≈20% faster) than YOLOv11, with near-identical mAP@0.5 across scenarios. B3 performed consistently well (mAP@0.5 ≈ 0.99), and class weighting improved its detection robustness without raising overall mAP. Residue detection remained the weakest across all scenarios (mAP@0.5 0.076–0.085) and worsened under weighting, indicating that ProgLoss and STAL alone do not resolve its structural visual heterogeneity; this weak, stable Residue performance was confirmed reproducible across three additional training runs with different seeds (mAP@0.5:0.95 = 0.046 ± 0.001). These findings partially support the research hypothesis, positioning YOLO26 (default) as a favorable accuracy-efficiency trade-off for automated waste sorting, while Residue detection requires further data enrichment and augmentation strategies beyond architectural improvements alone.]]>
Copyrights © 2026
