Denoised Temporal Relation Network for Temporal Action Segmentation

Zhichao Ma, Kan Li*

*此作品的通讯作者

科研成果: 书/报告/会议事项章节会议稿件同行评审

摘要

Temporal relations among action segments play a crucial role in temporal action segmentation. Existing methods tend to employ the graph neural network to model the temporal relation. However, the performance is unsatisfactory and exhibits serious over-segmentation due to the generated noisy features. To solve the above issues, we present an action segmentation framework, termed a denoised temporal relation network (DTRN). In DTRN, a temporal reasoning module (TRM) models inter-segment temporal relations and conducts feature denoising jointly. Specifically, the TRM conducts an uncertainty-gated reasoning mechanism for noise-immune and utilizes a cross-attention-based structure to combine the informative clues from the discriminative enhance module which is trained under Selective Margin Plasticity (SMP) to ensure informative clues, SMP adjusts the decision boundary adaptively by changing specific margins in real-time. Our framework is demonstrated to be effective and achieves state-of-the-art performance of accuracy, edit score, and F1 score on the challenging 50Salads, GTEA, and Breakfast benchmarks.

源语言英语
主期刊名Pattern Recognition and Computer Vision - 6th Chinese Conference, PRCV 2023, Proceedings
编辑Qingshan Liu, Hanzi Wang, Rongrong Ji, Zhanyu Ma, Weishi Zheng, Hongbin Zha, Xilin Chen, Liang Wang
出版商Springer Science and Business Media Deutschland GmbH
282-294
页数13
ISBN(印刷版)9789819985364
DOI
出版状态已出版 - 2024
活动6th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2023 - Xiamen, 中国
期限: 13 10月 202315 10月 2023

出版系列

姓名Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
14430 LNCS
ISSN(印刷版)0302-9743
ISSN(电子版)1611-3349

会议

会议6th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2023
国家/地区中国
Xiamen
时期13/10/2315/10/23

指纹

探究 'Denoised Temporal Relation Network for Temporal Action Segmentation' 的科研主题。它们共同构成独一无二的指纹。

引用此