Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ICCV 2019] ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning #7

Open
uhhyunjoo opened this issue Apr 13, 2022 · 1 comment
Labels
ICCV International Conference on Computer Vision V2V Video to Video Retrieval

Comments

@uhhyunjoo
Copy link
Owner

uhhyunjoo commented Apr 13, 2022

link
paper ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning
code papers with code
@uhhyunjoo uhhyunjoo added ICCV International Conference on Computer Vision V2V Video to Video Retrieval labels Apr 14, 2022
@uhhyunjoo uhhyunjoo changed the title new2 [ICCV 2019] ViSiL: Fine-grained Spatio-Temporal Video Similarity Learning Apr 14, 2022
@uhhyunjoo
Copy link
Owner Author

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
    1. 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
    2. 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
    3. train : a triplet loss scheme
  • 4개의 retreival 문제에 대해 5개의 벤치마크 데이터셋을 이용하여 evaluate 했고, sota 달성함

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
ICCV International Conference on Computer Vision V2V Video to Video Retrieval
Projects
Status: note 🔥
Development

No branches or pull requests

1 participant