Efficient annotation reduction with active learning for computer vision-based Retail Product Recognition

The retail industry encounters huge obstacles with computer vision (CV) technology due to frequent model retraining with changing products and time-consuming, costly data annotation. Previous research in this field has been primarily focused on optimizing model performance rather than minimizing annotation effort. Therefore, the main idea of this paper is to evaluate active learning as a method to minimize annotation effort in the retail industry. The MVTEC Densely Segmented Supermarket dataset is used to evaluate various active learning methods such as the Least Confident, Entropy and Cost-Effective Active Learning (CEAL) along with Mask R-CNN model. The results demonstrate that annotating only 20.83-24.34% of the data achieves 95% of the full dataset’s performance. When training, out-of-sample data share similar characteristics, the Least Confident and CEAL methods reduce annotation requirements by 7.7-15.7% while maintaining 95% and 97% of the full dataset’s performance. However, the Entropy method under-performs compared to the random selection baseline. Ultimately, none of the methods show a clear advantage when the data characteristics differ between training and out-of-sample data. Finally, the proposed active learning methods on an industry-specific retail dataset remarkably propels the development of highly efficient and cost-effective CV solutions meticulously tailored for the retail industry.