Presentation 2018-02-16
Style transfer for domain adaptation in object detection
Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, Kiyoharu Aizawa,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Domain adaptation is critical for success in new and unseen environments. Existing approaches for domain adaptation based on generative adversarial networks (GANs) produce images which resemble images in target domain from images in source domain. However, they sometimes fail to capture larger domain shifts. We propose a new image generation model based on style transfer. Our model explicitly extracts high-level and low-level features from one image from source and target domain, respectively, and synthesizes a new image. Extracting the features from each image in the target domain enables the model to create diverse and more similar images to the target domain. We test our model on novel datasets containing three image domains, and achieve an approximately 1 to 4 percent points improvement in terms of mean average precision (mAP) compared to the best-performing baselines.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) object detection / domain adaptation / style transfer
Paper # ITS2017-79,IE2017-111
Date of Issue 2018-02-08 (ITS, IE)

Conference Information
Committee ITS / IE / ITE-MMS / ITE-HI / ITE-ME / ITE-AIT
Conference Date 2018/2/15(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Hokkaido Univ.
Topics (in Japanese) (See Japanese page)
Topics (in English) Image Processing, etc.
Chair Takayoshi Yokota(Tottori Univ.) / Takayuki Hamamoto(Tokyo Univ. of Science) / Norihiko Ishii(NHK) / Masayuki Sato(Univ. of Kitakyushu) / Miki Haseyama(Hokkaido Univ.) / Nobuhiko Mukai(Tokyo Cisy Univ.)
Vice Chair Masahiro Fujii(Utsunomiya Univ.) / Tomotaka Wada(Kansai Univ.) / Kazuya Kodama(NII) / Hideaki Kimata(NTT) / / / Norio Tagawa(Tokyo Metropolitan Univ.) / Hisaki Nate(Tokyo Polytechnic Univ.)
Secretary Masahiro Fujii(Meiji Univ.) / Tomotaka Wada(Saitama Univ.) / Kazuya Kodama(Nagoya Univ.) / Hideaki Kimata(KDDI Research) / (NHK) / (NTT) / Norio Tagawa(NHK) / Hisaki Nate(NHK)
Assistant Msataka Imao(ME) / Yanlei Gu(Univ. of Tokyo) / Kenshi Saho(Toyama Prefectural Univ.) / Kouichiro Hashiura(Akita Prefectural Univ.) / Yasutaka Matsuo(NHK) / Kazuya Hayase(NTT)

Paper Information
Registration To Technical Committee on Intelligent Transport Systems Technology / Technical Committee on Image Engineering / Technical Group on Multi-media Storage / Technical Group on Human Inormation / Technical Group on Media Engineering / Technical Group on Artistic Image Technology
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Style transfer for domain adaptation in object detection
Sub Title (in English)
Keyword(1) object detection
Keyword(2) domain adaptation
Keyword(3) style transfer
1st Author's Name Naoto Inoue
1st Author's Affiliation The University of Tokyo(UTokyo)
2nd Author's Name Ryosuke Furuta
2nd Author's Affiliation The University of Tokyo(UTokyo)
3rd Author's Name Toshihiko Yamasaki
3rd Author's Affiliation The University of Tokyo(UTokyo)
4th Author's Name Kiyoharu Aizawa
4th Author's Affiliation The University of Tokyo(UTokyo)
Date 2018-02-16
Paper # ITS2017-79,IE2017-111
Volume (vol) vol.117
Number (no) ITS-431,IE-432
Page pp.pp.235-238(ITS), pp.235-238(IE),
#Pages 4
Date of Issue 2018-02-08 (ITS, IE)