TY - GEN
T1 - Virtual Try-On Networks Based on Images with Super Resolution Model
AU - Gallegos, Franco
AU - Contreras, Sebastian
AU - Ugarte, Willy
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2023
Y1 - 2023
N2 - The main job of a virtual imaging try-on is to transfer a garment to a specific area of an individual’s body part. Trying to deform said garment so that it fits in a part of the desired body. Despite some research, the vast majority use a low-quality image resolution of 192 × 256 pixels, limiting the visual satisfaction of online users. Analyzing this visual limitation, we find that the vast majority of the algorithms use these mentioned measures to obtain better performance and optimization during their training, since while the number of pixels is smaller, in the same way, their execution time will be less in the generation. of segments or masks of garments and body parts. Despite having better performance and optimization, quality and pixel size are also of the utmost importance, since it is in the final resolution that the result for the user is appreciated. To address this challenge, we propose a super-resolution extension module, added to the ACGPN model. Such a module gets the resulting image from the ACGPN model, and then with the help of computer vision aims to increase the resolution to 768 × 1024 pixels with minimal loss of quality. For this, a comparison of models that perform this task of increasing the resolution is made. Finally, it is quantitatively shown that the proposal obtains better results.
AB - The main job of a virtual imaging try-on is to transfer a garment to a specific area of an individual’s body part. Trying to deform said garment so that it fits in a part of the desired body. Despite some research, the vast majority use a low-quality image resolution of 192 × 256 pixels, limiting the visual satisfaction of online users. Analyzing this visual limitation, we find that the vast majority of the algorithms use these mentioned measures to obtain better performance and optimization during their training, since while the number of pixels is smaller, in the same way, their execution time will be less in the generation. of segments or masks of garments and body parts. Despite having better performance and optimization, quality and pixel size are also of the utmost importance, since it is in the final resolution that the result for the user is appreciated. To address this challenge, we propose a super-resolution extension module, added to the ACGPN model. Such a module gets the resulting image from the ACGPN model, and then with the help of computer vision aims to increase the resolution to 768 × 1024 pixels with minimal loss of quality. For this, a comparison of models that perform this task of increasing the resolution is made. Finally, it is quantitatively shown that the proposal obtains better results.
KW - Computer vision
KW - Generative network model
KW - Super resolution models
KW - Virtual try-on
UR - https://www.scopus.com/pages/publications/85180805454
U2 - 10.1007/978-3-031-49339-3_11
DO - 10.1007/978-3-031-49339-3_11
M3 - Contribución a la conferencia
AN - SCOPUS:85180805454
SN - 9783031493386
T3 - Communications in Computer and Information Science
SP - 182
EP - 194
BT - Innovative Intelligent Industrial Production and Logistics - 4th International Conference, IN4PL 2023, Proceedings
A2 - Terzi, Sergio
A2 - Madani, Kurosh
A2 - Gusikhin, Oleg
A2 - Panetto, Hervé
PB - Springer Science and Business Media Deutschland GmbH
T2 - 4th International Conference on Innovative Intelligent Industrial Production and Logistics, IN4PL 2023
Y2 - 15 November 2023 through 17 November 2023
ER -