Label quality as a bottleneck for YOLO detection in all-weather autonomous driving using the MUSES dataset.

Zreiqat, DZ; Nguyen, V; Rawashdeh, NA · 2026 · PubMed Central

DOI: 10.1007/s44291-026-00244-5

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study investigates the impact of annotation quality on the performance of YOLO-based object detection models in autonomous driving scenarios under adverse weather and low-illumination conditions. While modern detectors perform well in clear daytime settings, their accuracy degrades significantly in fog, rain, snow, and night. The authors argue that this performance gap is often attributed to model limitations or domain shift, but may be largely confounded by systematic label noise in training datasets. Using the forward-facing RGB subset of the MUSES dataset, the research quantifies how missing objects, mislocalized bounding boxes, and class inconsistencies limit the achievable accuracy of lightweight detectors, specifically YOLOv11n and YOLOv11s, which are chosen to reflect embedded ADAS constraints. To address this bottleneck, the authors developed a semi-automatic label-upgrading pipeline. The process begins with a baseline training on the original, noisy annotations, followed by a diagnostic audit that identified severe supervision issues, particularly for small, distant, or occluded objects. To correct these errors without full manual relabeling, a high-capacity teacher model (YOLOv12xL) generated candidate pseudo-labels for all images. These candidates were then verified and corrected by a human annotator, who refined bounding boxes, corrected class labels, removed false positives, and added missed but visible objects. The resulting cleaned dataset comprised 1,281 training, 215 validation, and 675 test images across eight weather-illumination conditions. The lightweight models were then retrained on this cleaned subset using identical hyperparameters and training protocols to isolate the effect of supervision quality. The results demonstrate that label quality is a primary bottleneck for all-weather detection. When trained on the original noisy annotations, both YOLOv11n and YOLOv11s achieved very low accuracy, with mAP@0.5 scores of approximately 0.13 and recall values below 0.14. Increasing model capacity from YOLOv11n to YOLOv11s yielded no measurable benefit, indicating that learning was limited by the training signal rather than representational power. However, retraining on the cleaned annotations produced substantial improvements: YOLOv11n mAP@0.5 increased from 0.130 to 0.557, and YOLOv11s increased from 0.127 to 0.620. Significant gains were also observed in mAP@0.5:0.95 and recall, confirming that improved supervision completeness and localization accuracy unlock large performance gains. The study concludes that supervision quality strongly influences the effectiveness of architectural improvements in adverse-weather settings. The findings suggest that apparent robustness failures in autonomous driving perception may be partly due to incomplete or inconsistent ground-truth labels rather than inherent model deficiencies. By providing a label-audited baseline, the paper highlights the importance of data-centric fixes and rigorous annotation auditing in benchmarking all-weather detection, showing that lightweight models can achieve strong accuracy once label noise is mitigated.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• weather rain fog snow