AdaptCD: An Adaptive Target Region-Based Commodity Detection System

Zeliang Ma, Delong Liu, Zhe Cui, Yanyun Zhao; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 5486-5495

With the rapid development of computer vision, the detection and counting of goods through computer vision techniques has become practicable. The AICity competition has focused its attention on the automatic recognition and counting of commodities, and has significantly propelled the advancement of this field through the organization of competitive events. Minimizing false positives and false negatives is critical to the success of this task. An Adaptive Target Region-based Commodity Detection System has been designed in this study to accurately identify the trajectory and category of goods. To alleviate the difference between training and testing data, two data augmentation methods are utilized, and various data synthesis methods are also designed to meet the training needs of different network models in the framework. Additional Adaptive Algorithms are designed to solve the problem of camera movement during product shooting. An Effective Fusion Algorithm is proposed for Dual Detectors to complement their advantages and minimize detection errors. To maximize the efficiency of the well-trained commodity classifier, an innovative Multi-layer Perception Fusion Module(MPFM) is devised to enhance the commodity classification capabilities, thereby generating more dependable features for tracking purposes. The system has been validated in Multi-Class Product Counting & Recognition for Automated Retail Checkout (2023 AI CITY Challenge Task4) competition, where our results achieve F1 Score of 0.9787 in Task4 testA, ranking second in 2023 AI CITY Challenge Task4. The code will be released at: https://github.com/mzl163/AICITY23_Task4