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

Large Scale Landmark Recognition via Deep Metric Learning #186

Open
chullhwan-song opened this issue Aug 30, 2019 · 1 comment
Open

Large Scale Landmark Recognition via Deep Metric Learning #186

chullhwan-song opened this issue Aug 30, 2019 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

https://arxiv.org/abs/1908.10192

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Sep 2, 2019

abstract

전체 프로세스

image

  • 특이(?)한점은 첫번째 과정에서, landmark vs non-landmark인것인가를 check
    • 이를 위해 , 각 클래스의 centroid를 계산 - 각 embedding feature의 average
      • 특이할점은, 하나의 클래스(같은 랜드마크)라도 촬영시점이 다른 것끼리는 분리
        image
  • base network > Wide Residual Network (WRN-50-2) > 잘 몰라서 더 찾아봐야할듯.
    image
  • feature > global average pooling
    image
  • loss
    • softmax+center loss
    • 논문제목이 metric learning이어서.. softmax ?? 를 사용한 케이스? center loss가 ?
      image
    • cam기반 landmark 설명한것을 보니 softmax는 확실 ㅎ
      image

학습

  • 제목그래로 metric learning인줄알았더니..softmax+center loss 조합
  • curriculum learning 개념으로 학습셋을 구성하여 학습
  • 기본적으로 이 회사에서 만든 학습셋인듯
    • (1) Europe (including all Russia);
    • (2) North America, Australia and Oceania;
    • (3) Middle East, North Africa;
    • (4) Far East.
      image

실험

  • delf 와 같이 통합하여 실험 - spatial verification
  • 다음부터 좀 헷갈린다.
  • train data ?
    • sensitivity — accuracy of a model on images with landmarks (also called Recall)
      • ?? 학습모델을 랜크마크 이미지를 가지고 구성하여 모델 생성
    • specificity — accuracy of a model on images without landmarks.
      • ?? 학습모델을 랜크마크 이미지없이 가지고 구성하여 모델 생성.
  • test data ?
    • “rare” and “frequent” landmarks.
      image
      • “Part from total number” : shows what percentage of landmark examples in the offline test
        has the corresponding type of landmarks > offline??몬지는 모르겠지만.ㅎ 전체 수집한것중 잘 나타나지 않는 landmark를 의미하는듯.??
        • 각 클래스를 보고 그 클래스의 갯수를 보고 자른것인가? 그렇다면 반대가 되어야하는데. ㅠ(유명 랜드마크는 그 클래스안에서는 개수는 많아도 전체 지구 개수의 랜드마크의 아주 일부분이기 때문에....?? 그래서 이 개념은 아닌듯) > 데이터를 주지 않으니..ㅠ
        • 특정 영역(geo)을 crawling를 했는데..그때의 전체 이미지중 보니 매우 낮은 랜드마크만 있었서 나머지는 non-landmark 비율이 높은 경우를 'rare'로 하지 않았을까?
          • It’s very important to understand how model works on “rare” landmarks due to the small amount of data for them.
  • 데이터셋구성
    • test set > all: 581545 > 이중 약 3%만 landmark 이미지
    • 아래 그림은 "rare" vs "frequent" > 그림보고도 이해안감.ㅠ > 대충? 설명
    • 아래 그림의 왼쪽은 centriod들과 밀접한 top 5 centriod 이미지 > 그래서, 각 클래스안에 다른 view가 존재하는듯..(위에서 언급)
      image
  • eval set 평가 = Medium vs Hard > distractors image에 대한 포함여부 테스트
    image
    • 평가가 살짝 fair하지 않는듯 > 여기서는 자체 train set를 ... > ( neuralcode같은것을 이용하지 않고..) > 공개했으면..
    • neuralcode나 google landmark 데이터셋을 trainset으로한 실험도 같이 했으면 더 좋을것 같다.

@chullhwan-song chullhwan-song reopened this Sep 5, 2019
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant