issues
search
chullhwan-song
/
Reading-Paper
151
stars
26
forks
source link
Feature Learning based Deep Supervised Hashing with Pairwise Labels
#79
Open
chullhwan-song
opened
5 years ago
chullhwan-song
commented
5 years ago
https://arxiv.org/abs/1511.03855
chullhwan-song
commented
5 years ago
Abstract
for large-scale image retrieval
hashing 기법 소개
deep pairwise-supervised hashing (DPSH)라는 알고리즘 제안
pairwise label > metric learning > hash-code learning
feature & hash-code learning을 동시에 학습
contribution
3개의 요소를 가진 end-to-end learning framework
to learn image representation from pixels
그냥 General한 CNN 구조를 가지는데..이미지를 처리..라서..?? ㅎ
image representation는 FC feature로 보이는듯. image -> conv -> fc
hash function to map the learned image representation to hash codes
image representation : 일종의 float vector로 추측되고 이를 hash code로 converion할 수 있는 hash function을 이용
a loss function to measure the quality of hash codes guided by the pairwise labels.
pairwise labels 즉, 한쌍의 라벨을 가진 hash code들(두개의 vector>hash)로 loss 측정
Siamese net기반 metric learning인듯
Notation and Problem Definition
본문에는 어려운데, 실제로는 간단?
해석해보면, vector를 (element-wise) sign function에 넣어 양수이면 1, 그렇지 않으면 -1
hamming distance에 적용가능한 binary code
learned
Model and Learning
pairwise label based supervised hashing methods
pairwise-supervised hashing (DPSH) : feature & hash-code learning을 동시에 학습
Fig.1에서 보듯이, Siamese net기반 metric learning
fc 이후에, hash function을 가진 layer를 두고 이를 이용하여 loss 계산.
자세한 network 구조는 다음 table을 보면
hash function 수식은
loss
pairwise loss function
negative log-likelihood > binary code로 distance를 구하는 형태 l2인듯..
Experiment
https://arxiv.org/abs/1511.03855