Rethinking Object Detection in Retail Stores

The conventional standard for object detection uses a bounding box to represent each individual instance. However, it is not practical in the industry-relevant applications context of warehouses due severe occlusions among groups instances same categories. In this paper, we propose new task, i.e., simultaneously localization and counting, abbreviated as Locount, which requires algorithms localize objects interest with number instances. there does exist dataset or benchmark designed such task. To end, collect large-scale counting rich annotations retail stores, consists 50,394 images more than 1.9 million 140 Together dataset, provide evaluation protocol divide training testing subsets fairly evaluate performance developing Locount Moreover, present cascaded network strong baseline, gradually classifies regresses boxes predicted numbers enclosed boxes, trained an end-to-end manner. Extensive experiments are conducted on proposed demonstrate its significance analysis provided indicate future directions. Dataset available at https://isrc.iscas.ac.cn/gitlab/research/locount-dataset.