An Effective Framework of Multi-Class Product Counting and Recognition for Automated Retail Checkout
As the field of computer vision grows, Automated Retail Checkout has become a highly anticipated development goal. The key of this task is to improve the accuracy rate. If there is an error, it will bring serious losses to the business and awful experience for customers which is not our expected. This competition gives us an opportunity to simulate check-out in a real world scenario, so that we can identify problems and solve them, not only for the competition, but also for the practical application. As one of the participating teams in this task, we pursue the goal of avoiding misdetection and misclassification, and build a complete set of framework to achieve high-precise, high-recall performance. In addition, there is an excessive difference between the training data and test data. How to use limited data to make up for the differences in this part is also one of the highlights of our framework. In general, our framework consists of three main parts. Firstly, the Pre-Processing module to make up for the differences between training and test data. The DTC module completes the overall process of automatic recognition. Finally the MTCR module is proposed to post-process the output of the DTC module. On the TestA data of AICITY2022 Task 4, we have achieved significant result compared to the other teams. Finally, our model is ranked 1st in AICITY2022 Task 4. The code is available at: https://github.com/w-sugar/DTC_AICITY2022.