search

ChaojianYu / Hierarchical-Bilinear-Pooling

Implementation for <Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition> in ECCV'18.
Other
103 stars 23 forks source link

input image size #6

Closed Blue-Clean closed 5 years ago

Blue-Clean commented 5 years ago

@ChaojianYu hi, I am confused about the input image size. I am new to this field. The input image size requires 448448 or not? I've downloaded the origin cub_200_2011 and resized them to 512S. Then I need to center crop them to 448448 as input image or directly take them 512S as input image size?

ChaojianYu commented 5 years ago

@Blue-Clean you should directly take 512*S image as input. The cropping operation will be set in the ImageData layer. For specific settings, you can refer to the ImageData layer of our prototxt file.

Blue-Clean commented 5 years ago

@ChaojianYu Thank you so much. Now I have some question about how to resize the image into 512S. I use python library PIL , new_image = img.resize((width,height),Image.BICUBIC).But the 512S image seems NOT high resolution compared with original one and the size of a certain image changes from 53.2kB into 25.1kB(compressed?). Can you tell me how to do? Forgive my ignorance.

ChaojianYu commented 5 years ago

@Blue-Clean It doesn't matter if the image is compressed. you can also download the preprocessed code here: https://drive.google.com/file/d/1BKrjLrcHmJHtSuUwDInL5uaozikJIaKD/view?usp=sharing

Blue-Clean commented 5 years ago

@ChaojianYu Thank you so much. HaHaHa, u r so nice!!