Adaptive Open Set Recognition with Multi-modal Joint Metric Learning

Yimin Fu, Zhunga Liu, Yanbo Yang, Linfeng Xu, Hua Lan

科研成果: 书/报告/会议事项章节会议稿件同行评审

1 引用 (Scopus)

摘要

Open set recognition (OSR) aims to simultaneously identify known classes and reject unknown classes. However, existing researches on open set recognition are usually based on single-modal data. Single-modal perception is susceptible to external interference, which may cause incorrect recognition. The multi-modal perception can be employed to improve the OSR performance thanks to the complementarity between different modalities. So we propose a new multi-modal open set recognition (MMOSR) method in this paper. The MMOSR network is constructed with joint metric learning in logit space. By doing this, it can avoid the feature representation gap between different modalities, and effectively estimate the decision boundaries. Moreover, the entropy-based adaptive weight fusion method is developed to combine the multi-modal perception information. The weights of different modalities are automatically determined according to the entropy in the logit space. A bigger entropy will lead to a smaller weight of the corresponding modality. This can effectively prevent the influence of disturbance. Scaling the fusion logits by the single-modal relative reachability further enhances the unknown detection ability. Experiments show that our method can achieve more robust open set recognition performance with multi-modal input compared with other methods.

源语言英语
主期刊名Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Proceedings
编辑Shiqi Yu, Jianguo Zhang, Zhaoxiang Zhang, Tieniu Tan, Pong C. Yuen, Yike Guo, Junwei Han, Jianhuang Lai
出版商Springer Science and Business Media Deutschland GmbH
631-644
页数14
ISBN(印刷版)9783031189067
DOI
出版状态已出版 - 2022
活动5th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2022 - Shenzhen, 中国
期限: 4 11月 20227 11月 2022

出版系列

姓名Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
13534 LNCS
ISSN(印刷版)0302-9743
ISSN(电子版)1611-3349

会议

会议5th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2022
国家/地区中国
Shenzhen
时期4/11/227/11/22

指纹

探究 'Adaptive Open Set Recognition with Multi-modal Joint Metric Learning' 的科研主题。它们共同构成独一无二的指纹。

引用此