Which Region Proposal to Choose? A Case Study for Automatic Identification of Retail Products

Identifying products visible in an image of a rack of a supermarket is a challenging and commercially relevant machine vision problem. For identification, the region proposal algorithm generates a number of (mostly overlapped) region proposals around each product on the rack. Each region proposal is then assigned a product class with a certain classification score. Finally, the products are detected using non-maximal suppression (NMS) discovering winners among the region pro-posals. Greedy-NMS takes classification scores of the proposals as a key factor and thereby often eliminates (geometrically) better-fitted proposals. Graph-based NMS (G-NMS) provides a better alternative performance-wise but an inferior solution time-complexity-wise FFFFFFFcompared to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$O(N^{2})$</tex> time-complexity of greedy-NMS. This paper introduces a adjusted classification score for use in the novel rectified non-maximal suppression (R-NMS) setup that runs in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$O(N^{2})$</tex>. The efficacy of the proposed adjusted classification score is theoretically characterized in better discriminating overlapped region proposals. Our experiments establish that the performance of the proposed R-NMS is never inferior to G-NMS, and outperforms greedy-NMS while testing on several datasets.