Triplet Network Architecture
Transcription
Triplet Network Architecture
Deep metric learning using Triplet network Elad Hoffer, Nir Ailon 발표자: 전혁준 요약 - Deep learning - - Triplet network model 제안 - - 유용한 의미의 표현을 학습 Classification task의 일부로 학습(기존) - black car, white car, dark grey car ⇒ car Similarity task(Fine-grained image similarity) - black car query ⇒ dark grey car > white car 거리 차로 표현을 학습 영상 검색에서 ranking을 학습에 사용(Wang et al.) 유사한 것을 더 가깝게, 다른 것은 더 멀게 Competitor - Siamese network Metric Learning - 유사/거리 함수 학습 K-NN, Clustering, Image Retrieval에 활용 http://horicky.blogspot.kr/2012/08/measuring-similarity-and-distance.html Face Recognition - Classification: 누구냐? - Metric Learning: 같냐? Same person or not. Siamese Network Architecture Learning Hierarchies of Invariant Features. Yann LeCun. helper.ipam.ucla.edu/publications/gss2012/gss2012_10739.pdf Triplet Network Architecture LOSS Layer F(x-) F(x) F(x+) Q P N x x+ x- Embedding Net Triplet Tuple에서 x-x1는 x-x2보다 더 유사하다. (Siamese Network에서는 표현을 못 함) ⇒ relative relationship Result Deep Ranking Mutiscale network structure. 3개의 path로 작동. 녹색 박스는 deep convolution network임. 마지막으로 l2 Normalization에서 정규화를 진행. Learning Fine-grained Image Similarity with Deep Ranking, Jiang Wang et al., https://arxiv.org/pdf/1404.4661.pdf Deep Ranking Result OASIS feature(L1HashKPCA feature with OASIS learning) Visual appearance는 잘 표현하지만, semantic은 떨어진다. ConvNet semantic 하지만, global 속성이 떨이진다. Deep Ranking은 visual appearance와 semantic 이 잘 된다.
Similar documents
Siamese Network - Robot Vision Group
closely coupled pair, so as to save space between them. http://en.wikipedia.org/wiki/Siamese
More information국제질서의변용과 영토문제
Chair : Paik Jin-Hyun (Judge, International Tribunal for the Law of the Sea) All participants
More information