unizard / AwesomeArxiv

89 stars 14 forks source link

[2017.11.01] Image Patch Matching Using Convolutional Descriptors with Euclidean Distance #37

Open unizard opened 6 years ago

unizard commented 6 years ago

ACCVW2016

Keywords: DeepDescriptor, Metric Learning, STN?

URL: https://arxiv.org/pdf/1710.11359.pdf Interest: 2

unizard commented 6 years ago

기존 피처의 경우 지역불변성을 가지기 위해, scale, rotation, localization에 대한 정규화 과정을 거친다. 생각해보니, shape에 대한 3D 볼륨 을 고려하지 않는 이유는, 전통적인 피쳐의 경우, 촬영된 피쳐는 plane위에 있다라는 가정을 가지고 간다.

전통적으로 3D Feature는 shape에 대한 normal 방향을 정규화 하여 기술하는 다양한 방법론들이다. 2.5D 라고 불리워 졌던 애들이 있었던거 같은데 기억은 안난다.

자자.. 그럼, 2D Feature에 shape 정보를 녹아 낼수 있는 정규화 방법론이 나오면, 의미가 있어질 것 같은데....ㅎ

QinZiwen commented 6 years ago

@unizard Have you had this paper's dataset(Image Patch Matching Using Convolutional Descriptors with Euclidean Distance)? Can you share me? Thank you very much!

unizard commented 6 years ago

@QinZiwen If you want to download Multi-view Stereo Correspondence Dataset (MSC), Please, see this webpage. http://matthewalunbrown.com/patchdata/patchdata.html