issues
search
uhhyunjoo
/
paper-notes
이슈로 가볍게 남깁니다.
0
stars
0
forks
source link
[ICCV 2019] ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning
#7
Open
uhhyunjoo
opened
2 years ago
uhhyunjoo
commented
2 years ago
link
paper
ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning
code
papers with code
uhhyunjoo
commented
2 years ago
Abstract
비디오 pairs 간의 fine-grained Spatio-Temporal relations 를 고려하는 Video Similarity Learning architure 인 ViSiL 을 제안함
해당 relations 는 이전의 video retrieval 방식에서는 고려되지 않았음
여기서 말하는 이전 방식은 whole frame 이나 whole video 를 embed 시켜서 a vector descriptor 로 만들고, 그 후에 similarity estimation 을 하는 것임
ViSiL 은 refined frame-to-frame similarity matrics 를 이용해서 video-to-video similarity 를 computation 을 학습하는 CNN-based 모델임
이로 인해 intra-frame relation 와 inter-frame relation 를 둘 다 고려 가능함
제안된 method
regional CNN features 에 Tensor Dot (TD) 랑 Chamfer Similarity (CS) 를 적용해서, pairwise frame similarity 를 estimate 함
이로 인해 frames 간의 similarity 가 계산되기 전에 feature aggregation 이 되는 것을 avoid 함
video frames 간의 similarity matrix 가 a four-layer CNN 에 fed 되고, 이후 CS 이용해서 a video-to-video similarity 로 만듦
이로 인해 videos 간의 similarity 가 계산되기 전에 featurea aggregation 이 되는 것을 avoid 함
이로 인해 matching frame sequences 간의 temporal similarity patterns 를 capture 함
train : a triplet loss scheme
4개의 retreival 문제에 대해 5개의 벤치마크 데이터셋을 이용하여 evaluate 했고, sota 달성함