Research Area:  Machine Learning
Given single instance (or template image) per product, our objective is to detect merchandise displayed in the images of racks available in a supermarket. Our end-to-end solution consists of three consecutive modules: exemplar-driven region proposal, classification followed by non-maximal suppression of the region proposals. The two-stage exemplar-driven region proposal works with the example or template of the product. The first stage estimates the scale between the template images of products and the rack image. The second stage generates proposals of potential regions using the estimated scale. Subsequently, the potential regions are classified using convolutional neural network. The generation and classification of region proposal do not need annotation of rack image in which products are recognized. In the end, the products are identified removing ambiguous overlapped region proposals using greedy non-maximal suppression. Extensive experiments are performed on one in-house dataset and three publicly available datasets: Grocery Products, WebMarket and GroZi-120. The proposed solution outperforms the competing approaches improving up to around 4% detection accuracy. Moreover, in the repeatability test, our solution is found to be better compared to state-of-the-art methods.
Keywords:  
Machine Vision System
Detection
Deep Learning
Machine Learning
Author(s) Name:  Bikash Santra, Avishek Kumar Shaw & Dipti Prasad Mukherjee
Journal name:  Machine Vision and Applications
Conferrence name:  
Publisher name:  Springer
DOI:  10.1007/s00138-021-01186-6
Volume Information:  volume 32, Article number: 56 (2021)
Paper Link:   https://link.springer.com/article/10.1007/s00138-021-01186-6