pris-cv / fine-grained-or-not Goto Github PK

View Code? Open in Web Editor NEW

56.0 4.0 9.0 14.84 MB

Code release for Your “Flamingo” is My “Bird”: Fine-Grained, or Not (CVPR 2021 Oral)

License: MIT License

Python 100.00%

fine-grained-classification fine-grained-visual-categorization fine-grained hierarchy-analysis multi-task-learning

fine-grained-or-not's Introduction

Fine-Grained-or-Not

Code release for Your “Flamingo” is My “Bird”: Fine-Grained, or Not (CVPR 2021 Oral) DOI

Changelog

2021/03/05 upload the code.

Requirements

python 3.6
PyTorch 1.2.0
torchvision

Data

Download datasets
Extract them to data/cars/, data/birds/ and data/airs/, respectively.
Split the dataset into train and test folder, the index of each class should follow the Birds.xls, Air.xls, and Cars.xls

e.g., CUB-200-2011 dataset

  -/birds/train
	         └─── 001.Black_footed_Albatross
	                   └─── Black_Footed_Albatross_0001_796111.jpg
	                   └─── ...
	         └─── 002.Laysan_Albatross
	         └─── 003.Sooty_Albatross
	         └─── ...
   -/birds/test	
             └─── ...

Training

python Birds_ours_resnet.py or python Air_ours_resnet.py or python Cars_ours_resnet.py

Citation

If you find this paper useful in your research, please consider citing:

@InProceedings{Chang2021Labrador,
  title={Your “Flamingo” is My “Bird”: Fine-Grained, or Not},
  author={Chang, Dongliang and Pang, Kaiyue and Zheng, Yixiao and Ma, Zhanyu and Song, Yi-Zhe and Guo, Jun},
  booktitle = {Computer Vision and Pattern Recognition},
  year={2021}
}

Contact

Thanks for your attention! If you have any suggestion or question, you can leave a message here or contact us directly:

fine-grained-or-not's People

Contributors

Stargazers

Watchers

Forkers

xuweigogogo hongbo-sun killsking eaststarkim roy881020 mengzmd lufei-github ilovesea xuzhikangnba

fine-grained-or-not's Issues

AIR和CAR数据集

您好，非常感谢您可以分享代码，但是我在复现您论文中的实验结果的过程中发现AIR数据集和CAR数据集与论文所展示的结果相差特别大，是否air.xls和car,xls文件有错？
希望得到您的回复，再次感谢！

Selection of some parameters and loss

hi，this works very well, I have two questions

Why Hiden_Number is 600？
Why choose NLL loss，do not use CE loss directly？

How to evaluate the performance under the traditional FGVC setting?

Hi,

Thanks for your amazing code and work!

What is the difference between the ACC of traditional FGVC setting and specie_acc? And how to change the code to evaluate the performance of the model under the traditional FGVC setting?

Training set of FGVC-Aircraft

Split the dataset into train and test folder, the index of each class should follow the Birds.xls, Air.xls, and Cars.xls

Hi, did you merge the training and validation set of FGVC-Aircraft into a new training set in your settings? Thank you!

how to reach 94.3 accuracy on stanford-cars？

hi！When the image size is 224224, the accuracy rate can reach about 89.4（species_acc）. When I change the image size to 448448, the accuracy rate is less than 80. How to achieve the 94.3 accuracy rate in the paper? Have you modified any other training parameters（epoch or lr or batchsize or others）?

A wrong label in Birds_name.xlsx for CUB-200-2011

Hi, I was using the Birds_name.xlsx file to build a tree of labels when I noticed there's an error in the file at row 85:

species | family | order
Red_legged_Kittiwake | Laridae | Coraciiformes

the order of Red_legged_Kittiwake should be Charadriiformes

Loss is nan？

hello,why does the training loss keep increasing and gradually become nan？
(python 3.8 torch 1.7 torchvision 0.8)

Test set of CUB

Greetings!
Love this work
But I have a silly question, there are some strange species in sheet2 in birds.xlsx, what should I do? Shoud I make some dir whose name is that ?

Cross Entropy loss or NLL loss

Hi, Thanks for the nice repo!

I just wanted clarification about the NLL criterion in the code. Are the results in Birds.out using Cross Entropy criterion or NLL criterion?

Also, I tried running the code as is (using NLL criterion), train and test loss becomes nan.

Thanks once again.

about convergence speed of training process

Hello, I tried your method on CUB Bird but found the printed loss is different from your log in this repo. Is there any details have been changed? Thank you.

The name list of family and species?

Greetings!
Love this work. Could you release the name list of family and species, please? I am curious about how you organize the 200 bird names.