fgvc

CAL-FGVC

This folder contains the implementation of the fine-grained image classification experiments.

Our implementation is based on the Pytorch version code of WS-DAN.

Prepare the data

CUB

Download CUB-200-2011 dataset from this link and move the uncompressed data folder to ./CUB-200-2011. The data structure should be:

./CUB-200-2011
        └─── images.txt
        └─── image_class_labels.txt
        └─── train_test_split.txt
        └─── images
                └─── 001.Black_footed_Albatross
                        └─── Black_Footed_Albatross_0001_796111.jpg
                        └─── ...
                └─── 002.Laysan_Albatross
                └─── ...

Stanford Cars

Download Stanford Cars dataset from this link and move the uncompressed data folder to ./stanford_cars. The data structure should be:

-/stanford_cars
      └─── cars_test
                └─── 00001.jpg
                └─── 00002.jpg
                └─── ...
      └─── cars_train
                └─── 00001.jpg
                └─── 00002.jpg
                └─── ...
      └─── devkit
                └─── cars_train_annos.mat
      └─── cars_test_annos_withlabels.mat

FGVC-Aircraft

Download FGVC-Aircraft dataset from this like and move the uncompressed data folder to ./fgvc-aircraft-2013b. The data structure should be:

./fgvc-aircraft-2013b/data/
                └─── images
                        └─── 0034309.jpg
                        └─── 0034958.jpg
                        └─── ...
                └─── variants.txt
                └─── images_variant_trainval.txt
                └─── images_variant_test.txt

Training & Evaluation

• Modify config_distributed.py to run experiments on different datasets
• Run bash train_distributed.sh to train models.
• Set configurations in config_infer.py and run python infer.py to conduct multi-crop evaluation.

Requirements

• Python 3
• PyTorch 1.0+
• Apex