Neural Network Enhanced Automatic Garment Measurement System Pawel Kowaleczko *1,2 , Przemyslaw Rokita 1 , Marcin Szczuka 3,4 1 Institute of Computer Science, Warsaw University of Technology ul. Nowowiejska 15/19, 00-665 Warsaw, Poland 2 QED Software sp. z o.o. ul.Miedziana 3A/18, 00-814 Warsaw, Poland 3 Institute of Informatics, University of Warsaw Banacha 2, 02-097 Warsaw, Poland 3 BAKERS sp. z o.o. ul. Branickiego 11/154, 02-972 Warsaw, Poland Email: pawel.kowaleczko.dokt@pw.edu.pl, pro@ii.pw.edu.pl, szczuka@mimuw.edu.pl Abstract—The measurement of garments is most often a very laborious task. Automatic garment measurement systems may be thus a great convenience in fashion e-commerce cataloguing issues. In this paper, we propose an automatic garment mea- surement system that uses classical computer vision algorithms, as well as an error correction neural network, which reduces the overall error. We make use of data collected by our partner, which contains photographs of garments with ArUco markers. Using such data, we estimate the coordinates of feature points, which are used to calculate a specific size of a garment. We apply the error correction neural network to this measured size to minimize the error. The conducted experiments show, that our method is a useful tool that meets the requirements of practicality and its results are comparable with the current state of the art methods. Additionally, our error correction neural network is a novelty in the field of automatic garment measurement and there is no need for the garments templates, which are used in the previous solutions. I. I NTRODUCTION I N RECENT years, a huge increase of interest in online shopping can be observed. Since 2017, the global e- commerce market almost doubled its value [1]. Because of its convenience, more and more people do their shopping online. However, shopping for garments may very often be problematic because of the inconsistency in sizes grading. The same piece of cloth may be tagged as size S by one producer, and as size M by another. This inconsistency is the cause of many returns of orders, which is one of the main problems and unnecessary expenses of running a fashion retail business online. To reduce the risk of return it is beneficial to additionally provide the specific dimensions, like sleeve length, waist width etc., in one of the common metrics - centimetres or inches. The accurate measurement technique is particularly useful in the case of the online second-hand shops. The clothes sold This research was co-funded by Smart Growth Operational Programme 2014-2020, financed by European Regional Development Fund, in frame of project POIR.01.01.01-00-0420/20, operated by National Centre for Research and Development in Poland. * Corresponding author there come from different suppliers and producers, whose size tables may be varied and inconsistent. In some cases, the clothes may have no indication of size whatsoever. The solutions proposed in this paper are an answer to real-life challenges encountered during the commercial R&D project. The Kidihub (kidihub.com) online second-hand shop specialises in gathering, listing and selling used children’s clothes. In Kidihub’s business model the clothes are being sent to the company’s warehouse by parents, and then assessed and qualified by the staff. The assessment process not only involves verification of quality and state (new, like new, small defect, ...) of the particular piece of clothing, but also leads to more extensive labelling of items. This labelling (tagging) is indeed the main focus of the R&D project. Once a piece of clothing is qualified as worthy listing, it is being ironed and arranged for picture taking. Then, the items need to be tagged with a number of feature values such as sex, colour, pattern, type, and – most crucially – size. The strive to increase the automation of measurement process and reduce the human involvement in the arrangement of clothes and picture-taking is what drives this research. We propose a method for automatic garment measurement which uses as an input a photograph of a piece of cloth. This photograph includes ArUco marker [2] for the sake of the need to determine the pixel to centimetre ratio to calculate the dimensions in centimetres. However, any marker, even a piece of cardboard, can be used for this task if the detection algorithm of this marker is implemented. We extract the contours of the garment and find the key feature points essential to determine the specific dimensions, such as points where sleeves end, hips edge points of pants etc. We calculate the Euclidean distance between those points and correct the error using a neural network. The main task of this network is to reduce the calibration error - the photos were taken in different camera to plane distances setups and only a single ArUco marker was provided, so the calibration was not possible. This neural network is also the main novelty of our work, one that makes it possible to reduce the error Communication Papers of the of the 17 th Conference on Computer Science and Intelligence Systems pp. 33–38 DOI: 10.15439/2022F8 ISSN 2300-5963 ACSIS, Vol. 32 ©2022, PTI 33