Naïve Approach for Bounding Box Annotation and Object Detection Towards Smart Retail Systems

摘要

It is becoming a trend that companies use smart retail stores to reduce the selling cost, by using the sensor technologies. Deep con-volutional neural network models which are pre-rained for the Object, detection task achieve state-of-the-art result in many benchmark. However, when applying these algorithms to the intelligent retail system to help automated checkout, we need to reduce the manual labelling cost of making retail data sets, and to achieve real-time demand while ensuring accuracy. In our paper, we propose a naive approach to get first portion of the bounding box annotations for a given custom image dataset in order to reduce manual cost. Experimental results show that our approach helps to label the first set of images in short time of period. Further, the custom module we designed helped to reduce the number of parameters by 41.77% for the YOLO model maintaining the original model's accuracy (85.8 mAP).