RetailDet: An Efficient Fusion Attention Network for Joint Product and Vacancy Identification in Smart Retail

Dense retail scenes pose unique challenges for both product detection and vacancy identification, where items are tightly packed on shelves and spaces between products carry critical inventory management implications. However, the existing detection frameworks, including state-of-the-art (SOTA) YOLO models, face limitations in feature fusion, particularly when integrating multi-level features and interpreting complex shelf arrangements. To address these challenges, we propose the RetailDet detection architecture and establish the RPV1K retail scene dataset. Our main contributions include: 1) A multi-attention module architecture that effectively fuses RGB and depth information for both product and vacancy detection; 2) A novel implicit gradient regulation (IGR) mechanism is proposed to address the redundant decision modules and feature modularization problems in traditional detection models. The proposed IGR mechanism dynamically regulates the gradient flow, optimizing the feature fusion path only during the training phase and not participating in calculations during the inference phase, thereby reducing computational overhead and enhancing model generalization performance; and 3) the RPV1K dataset, which is specifically tailored for retail detection tasks featuring product-vacancy co-occurrence scenarios. Experimental results show RetailDet significantly outperforms existing methods on the RPV1K dataset. Compared with YOLOv11-nano, RetailDet-nano achieves 10.6% mAP improvement while reducing inference latency by 17.2 ms. RetailDet-large achieves 85.2% precision with 20.79 M parameters. These innovations provide strong technical support for automatic shelf management and inventory replenishment in smart retail environments.