issues
search
chullhwan-song
/
Reading-Paper
151
stars
26
forks
source link
Neural Codes for Image Retrieval
#14
Open
chullhwan-song
opened
6 years ago
chullhwan-song
commented
6 years ago
https://arxiv.org/abs/1404.1777
chullhwan-song
commented
6 years ago
what?
neural code
neural code만의 새로운 landmark 데이터셋 공개
we chose a (semi)-automated approach for that.
by selecting 10,000 most viewed landmark Wikipedia pages (over the last month).
we used the title of the page as a query to Yandex image search engine.
then downloaded 1000 top images returned in response to the query (or less, if the query returned less images).
한마디로, text기반의 검색으로 이미지를 수집
따라서, 노이즈가 좀 많음. > triplet r-mac(DIR)에서는 이런 데이터 특성으로 인해 noisy를 제거하여 학습셋으로 사용. 덧붙혀, google delf 도 이를 이용
최종적으로, 672 classes and 213,678 images.
http://sites.skoltech.ru/compvision/projects/neuralcodes
imagenet 기반의 학습 모델을 이용. 그리고 위의 landmark set을 학습셋으로한 fine-tuning한 deep feature를 적용된 초기 논문
alexnet > "ImageNet Classification with Deep Convolutional Neural Networks"
deep feature : fully-connected layers 를 이용한다고 보이지만, 위의 그림에서는 conv feature map도 적용한 모습이 보임
최종적으로 PCA를 적용한 feature
실험
full descriptor
compression
imagenet model vs fine-tuning ?
https://arxiv.org/abs/1404.1777