敵対的性質を示す仮想類似データを利用した深層距離学習の汎化性能の向上

• 松岡 佑磨

書誌事項

タイトル別名

• VIRTUAL ADVERSARIAL SIMILAR POINT TO IMPROVE GENERALIZATION OF DEEP METRIC LEARNING

抄録

type:Article

Deep Metric Learning learns a small dimensional feature representation from input data points which has a geometry same as the input data points, in where the distance between similar data points are small and the distance between dissimilar datapoint are large. Therefore, it has been widely used in a variety of tasks like image retrieval and person re-identification. However, it requires to sample some kind of input data points to calculate similarity and dissimilarity to optimize itself but getting difficult to find efficient variations by hard example mining during its training. In this paper, we propose a novel deep metric learning method which is optimized by a loss function with generated virtual adversarial similar point and a metric loss and evaluate its performance in the Zero-shot learning benchmark with CUB-200-2011 and CARS-198 datasets.

収録刊行物

• 法政大学大学院紀要. 理工学・工学研究科編

  法政大学大学院紀要. 理工学・工学研究科編 60 1-6, 2019-03-31

  法政大学大学院理工学研究科

キーワード

• deep metric learning
• adversarial examples
• fine-grained image recognition
• zero-shot learning

詳細情報

• CRID: 1390290699801668352
• NII論文ID: 120006715031
• NII書誌ID: AA12677220
• ISSN: 21879923
• DOI: 10.15002/00022055
• Web Site: http://hdl.handle.net/10114/00022055
• 本文言語コード: ja
• データソース種別:
  • JaLC
  • IRDB
  • CiNii Articles