dataset

Dataset Curation for LT-DS Benchmarks

Data Folder Structure

DATA
├── awa2-lts
│   ├── image_list
│   │   ├── Original_train.txt
│   │   ├── Original_val.txt
│   │   ├── Original_test.txt
│   │   ├── Hayao_train.txt
│   │   ...    
│   ├── embedding
│   │   ├── ...
│   ├── original
│   │   ├── antelope
│   │   │   ├── antelope_00001.jpg
│   │   │   ├── antelope_00002.jpg
│   │   │   ...
│   │   ├── bat
│   │   ├── ...
│   ├── hayao
│   │   ├── antelope
│   │   │   ├── antelope_00001_hayao.jpg
│   │   │   ├── antelope_00002_hayao.jpg
│   │   │   ...
│   │   ...
│   ├── shinkai
│   │   ├── antelope
│   │   ...
│   ├── ukiyoe
│   │   ├── antelope
│   │   ...
│   └── vangogh
│       ├── antelope
│       ...
└── imagenet-lts
    ├── image_list
    │   ├── ...
    ├── embedding
    │   ├── ...
    ├── original
    │   ├── ILSVRC2012_val_00048700.JPEG
    │   ...
    ├── hayao
    │   ├── ILSVRC2012_val_00048700_hayao.jpg    
    │   ...
    ├── shinkai
    │   ├── ...
    ├── ukiyoe
    │   ├── ...
    └── vangogh
        ├── ...
        

Download the datasets

Firstly, download the original AWA2 and ImageNet datasets. And put it to the DATA/ structured as above

• AWA2

  1. Download from here.
  2. Put it to the DATA/awa2/original

• ImageNet-LT

  1. Download ImageNet_2014, and the ImageNet-LT file list.
  2. Based on the filelist, extract images from ImageNet_LT_train.txt, ImageNet_LT_test.txt, ImageNet_LT_val.txt, and put them to DATA/imagenet/original

Style Transfer

We use CartoonGAN for Hayao and Shinkai styles, and CycleGAN for Vangogh and Ukiyoe styles

Given the original sample with the name X.jpg, its corresponding style-transferred images are named as X_hayao.jpg, X_vangogh.jpg, X_shinkai.jpg, X_ukiyoe.jpg

Steps

Change REPO with this repo path, DATA with DATA folder path

Hayao and Sinkai

git clone https://github.com/Yijunmaverick/CartoonGAN-Test-Pytorch-Torch
cd CartoonGAN-Test-Pytorch-Torch
source pretrained_model/download_pth.sh
cp REPO/dataset/transfer_hs.py .
python transfer_hs.py --dataset awa2 --datapath DATA --style Hayao

Vangogh and Ukiyoe

git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
cd pytorch-CycleGAN-and-pix2pix
source ./scripts/download_cyclegan_model.sh style_vangogh 
source ./scripts/download_cyclegan_model.sh style_ukiyoe 
cp REPO/dataset/transfer_vu.py .
python transfer_vu.py --dataset imagenet --datapath DATA --style vangogh --dataroot . --name style_vangogh --model test

Train/Val/Test Splits

We provide the pre-defined train/validation/test index files to benchmark related algorithms.

Download from here and put to corresponding sub-folders under DATA/

Semantic Embeddings

Download the embeddings utilized in our experiments. Download from here and put to corresponding sub-folders under DATA/

Acknowledgement

CartoonGAN, CycleGAN, ImageNet-LT, AWA2