<![CDATA[On-Shelf Availability (OSA) is a critical aspect of retail operations that affects customer satisfaction and potential sales. Computer vision–based systems have emerged as a promising solution to monitor product availability on store shelves. However, their implementation faces the challenge of limited labeled data, which requires time-consuming manual annotation with precise bounding boxes. This study proposes a semi-supervised learning approach based on pseudo-labeling using the YOLOv11n architecture to address the scarcity of labeled data. We utilized a dataset of 918 retail product images with 174 classes, divided into four proportions of labeled data (20%, 40%, 60%, and 80%). The research stages included training a teacher model, generating pseudo-labels with a confidence threshold of 0.5, and training a student model using a combination of labeled and pseudo-labeled data. Experimental results show that this approach effectively improves detection performance. With 60% labeled data, the model achieved an mAP50 of 0.931 and an mAP50-95 of 0.864, along with high-quality pseudo-labels (F1-Score 0.727; IoU 0.819). This significant improvement indicates that pseudo-labels can enrich data variation without introducing excessive noise. The study demonstrates that semi-supervised learning can reduce dependence on large labeled datasets while offering a practical and efficient solution for OSA detection systems in retail environments]]>
Copyrights © 2025
