search

JDAI-CV / DCL

Destruction and Construction Learning for Fine-grained Image Recognition
Other
585 stars 159 forks source link

ct_train/val/test file #11

Closed yangyangkiki closed 5 years ago

yangyangkiki commented 5 years ago

Hi,

Can you share the ct_train/val/test.txt files of CUB/ Stanford Car/Aircraft to us?

Thank you.

Kind regards.

Kiki

Chen94yue commented 5 years ago

I added these files, although I think it will only take you a little time to generate them.

yangyangkiki commented 5 years ago

ok, thx a lot.

yangyangkiki commented 5 years ago

Hi, I download CUB on this web: http://www.vision.caltech.edu/visipedia/CUB-200.html But I found some images in the train.txt that provided by you are not exist. I don't know which part I did is wrong. Look forward to you reply. Thanks

Chen94yue commented 5 years ago

The right address for CUB-200-2011 dataset is: http://www.vision.caltech.edu/visipedia/CUB-200-2011.html Your address is the homepage for CUB-200-2010 dataset

yangyangkiki commented 5 years ago

ok, thank you.

yangyangkiki commented 5 years ago

@Chen94yue hi, should label = self.labels[item] (line 84) in dataset_DCL.py be revised to label = self.labels[item] -1 ?

Chen94yue commented 5 years ago

You are right. Sorry, I forget the code has been changed.

yangyangkiki commented 5 years ago

ok.

yangyangkiki commented 5 years ago

Hi @Chen94yue ，

I am confused that why you do this: img = img.crop((10, 10, widthcut-10, highcut-10)) https://github.com/JDAI-CV/DCL/blob/master/transforms/functional.py#L377 in transforms/functional.py when you swap the image.

So why you remove the edged 10 pixels from the image? Can you explain this? If I do not do this, I think it is better to find the similar patch through the mean value calculated before.

Thank you.

Chen94yue commented 5 years ago

The images with border are common in CUB images. This operation cuts the border of the images and is useless in the latest version. If your training images don't have border, you can comment out this line.

yangyangkiki commented 5 years ago

Got it. Thank you.