Interpretable workwear detection improves accuracy in low light conditions
DICL-Qwen3VL: Decoupled Confidence Calibration for Interpretable Workwear Detection
Computer Vision and Pattern Recognition
Summary
In indoor construction sites, dim lighting and blocked views make it hard to tell if people are wearing the right safety clothes. The authors designed a system that can spot upper-body workwear and explain its decisions with separate confidence scores for what it sees and how sure it is about rule compliance. This helps supervisors better understand when the system is uncertain because of the image or because of the rules. The system was tested on two datasets, showing better detection than earlier methods and providing clear evidence to support safety monitoring.
What this means in practice
- •For construction safety teams: Use interpretable workwear detection to verify safety compliance with clear confidence indicators in indoor construction environments.
- •For industrial inspection teams: Monitor worker apparel in factories with improved detection and interpretable evidence even under poor lighting or occlusion.
Authors
Meng Wang, Wu Zeng, Mei Li, Chao He
Abstract
Poor illumination and occlusion obscure clothing cues in indoor work sites, while diverse clothing styles complicate workwear compliance judgments. Although vision-language models (VLMs) support localization and semantic reasoning, their predictions often lack explicit supporting evidence, and a single confidence score cannot distinguish visual uncertainty from ambiguity in compliance judgments. We propose DICL-Qwen3VL, a decoupled interpretable VLM that integrates structured workwear detection, evidence generation, and separate visual and reasoning confidence estimates. First, low-rank adaptation (LoRA) adapts Qwen3-VL to localize visible upper-body clothing and identify workwear. Second, decoupled calibration separates visual observations from compliance reasoning and assigns a confidence score to each. Token entropy and sensitivity to image perturbations provide supervision for visual confidence, linking confidence estimation to the generated observations and their dependence on image content. Finally, group relative policy optimization (GRPO) jointly optimizes detection accuracy, output validity, and confidence calibration. Experiments on the company-collected Indoor Construction Site Workwear Dataset (ICSWD) and the web-collected Indoor Construction Workwear Dataset (WICWD) yield mAP@50 values of 0.730 and 0.753, respectively, outperforming the evaluated LoRA-adapted VLM baselines, including Qwen3-VL. The framework also provides explicit evidence and separate confidence scores to support manual review in industrial safety monitoring. WICWD is publicly available at https://github.com/DaErwen1/WICWD.