Workwear compliance checks in indoor construction sites are complicated by poor lighting, occlusion, and varied clothing styles. The authors argue that vision-language models (VLMs) can localize and reason about clothing, but their predictions often lack explicit supporting evidence, and a single confidence score cannot tell apart visual uncertainty from ambiguity in the compliance judgment itself.

To address this, the paper proposes DICL-Qwen3-VL, which adapts Qwen3-VL with low-rank adaptation (LoRA) to detect visible upper-body clothing and identify workwear. The framework then decouples calibration into separate visual and reasoning confidence scores. Visual confidence is supervised using token entropy and sensitivity to image perturbations, linking the confidence estimate to the generated observations and their dependence on image content. 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) report 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.