Geometry-Supervised Pose Network for Accurate Retail Shelf Pose Estimation

Abstract

In the smart retail industry, the quality of image collection is known to heavily affect the final analysis results (like accuracy) of applications, such as commodity detection, identification, and stitching. In practice, images captured manually by a monocular camera like mobile phone contain many low-quality images caused by an irregular shoot step. After image collection, filtering out low-quality images is a key step to mitigate the aforementioned impacts. One of the most effective solutions is to filter images with huge off-angles in 3-D through the shelf pose estimation algorithm. However, most of the existing camera pose estimation algorithms are designed for natural scenes and are difficult to realize in the structured real target scenes (like shelf scenes). Meanwhile, due to the lack of shelf pose dataset in academia and industry, there is still no approach designed for the shelf pose estimation in the smart retail scenario. In this article, we try to regress the complete shelf pose within a single end-to-end network and propose a novel geometry-supervised pose network (GSPN), which supervises the shelf pose estimation by learning the intrinsically geometric properties of shelves. Furthermore, we introduce the first retail shelf pose dataset (RSPD), including 28 876 images selected from three different shelf categories and being annotated carefully, as well as a complete 3-D shelf posture. The whole networks can be trained end to end with the shelf images and well-annotated ground truth. Experiments result of five strategies show that GSPN achieves the state-of-the-art performance on RSPD.