ETRI-Knowledge Sharing Plaform

ENGLISH

성과물

논문 검색
구분 SCI
연도 ~ 키워드

상세정보

학술대회 Various-Level Spatio-Temporal Alignment for Cross-Domain Action Recognition
Cited 0 time in scopus Download 2 time Share share facebook twitter linkedin kakaostory
저자
김형민, 김도형, 김재홍
발행일
202112
출처
International Conference on Robot Intelligence Technology and Applications (RITA) 2021 (LNNS 429), v.429, pp.323-335
DOI
https://dx.doi.org/10.1007/978-3-030-97672-9_29
협약과제
21HS1500, 고령 사회에 대응하기 위한 실환경 휴먼케어 로봇 기술 개발, 이재연
초록
Cross-domain action recognition is a less explored field of research until recently. The previous approaches usually feed the pre-extracted video-level or segment-level feature vectors to shallow networks for regenerating them. These approaches cannot directly affect the full capability of the action recognition models. Considering the recent researches, CNN has an inductive bias towards the texture of images. In a domain-changing situations, this information is affected and changed easily. Moreover theses low-level information is mainly encoded in intermediate features and is also needed to be aligned. For exploring the effect of the adaptation between various levels of spatial dimensions of the feature map, we divided the model into several parts and performed adaptation for each step. However, not every stage play the important role in action recognition in the temporal axis. To more sensitive adaptation, we propose a similarity-based weighting strategy. We first calculate the discrimination loss for each stage. Next, these discrimination losses are weighted by their similarity. The discrimination losses become large if the source and target sample's similarity values are small in a certain stage. The proposed method achieves state-of-the-art performance on UCF101-HMDB full dataset.
KSP 제안 키워드
Action recognition, Art performance, Cross Domain, Feature Map, Feature Vector, Inductive bias, Recognition model, Similarity-based, Weighting strategy, shallow networks, similarity values