講演抄録/キーワード |
講演名 |
2021-12-16 14:45
Unsupervised Logo Detection Using Adversarial Learning from Synthetic to Real Images ○Rahul Kumar Jain(Ritsumeikan Univ.)・Takahiro Sato・Taro Watasue・Tomohiro Nakagawa(tiwaki)・Yutaro Iwamoto(Ritsumeikan Univ.)・Xiang Ruan(tiwaki)・Yen-Wei Chen(Ritsumeikan Univ.) PRMU2021-31 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
Most of the existing deep learning based logo detection methods typically use a large amount of annotated training data, assuming that the training and test data belong to the same data distribution. Synthesized training images with automatically generated object-level annotations can be a solution to avoid the labor-intensive and time-consuming object annotation task. However, real-world problems limit this assumption and object detectors face domain-shift problems resulting in performance degradation. Here, we address the domain-shift problem in the field of logo detection from synthetic to real images. In this paper, to align the domain gap from synthetic to real image, we propose to use entropy minimization of mid-level output feature maps. We also synthesize training images using various data augmentation methods to perform experiments. Our experiments show that our proposed method improves performance by around 4% mAP compared to direct transfer from source to target domain (synthetic-to-real images) without any labeling cost and increasing network parameters. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Unsupervised Domain Adaptation / Adversarial Learning / Anchorless Object Detectors / Entropy Minimization / / / / |
文献情報 |
信学技報, vol. 121, no. 304, PRMU2021-31, pp. 43-44, 2021年12月. |
資料番号 |
PRMU2021-31 |
発行日 |
2021-12-09 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2021-31 |