README.md

Data

Overall structure of this directory should be as follows.

data
   ├── coco
   ├── imgnet
   ├── CIRR
   └── fashion-iq

ImageNet

 imgnet
    ├── imagenet-r ## unzipped imagenet-r directories containing images. This folder should contain subfolders.
            └──n01443537
                  .
                  . 

    ├── imgnet_real_query.txt
    ├── imgnet_targets.txt
    └── real ## imagenet validation directories containing images. This folder should contain subfolders.
        └──n01440764
                  .
                  . 

See ImageNet-R to download the dataset.

COCO

 coco
   ├── annotations/instances_val2017.json ## annotations for COCO validation images.
   ├── prepare_data.py ## code to generate query data.
   ├── coco_eval.csv ## this will be generated by running prepare_data.py
   ├── val2017 ## directory containing COCO validation images.
   └── val2017_masked ## running prepare_data.py will produce the directory.

Download both instances_val2017.json and val2017. Run the command to below to produce directory of val2017_masked.

python prepare_data.py

CIRR

cirr
  ├── captions
        └──cap.rc2.val.json
  ├── dev
  └── image_splits
        └──split.rc2.val.json

Download the images following instruction on CIRR.

Fashion-IQ

fashion-iq
    ├── json
        ├── cap.dress.val.json
        ├── cap.shirt.val.json
        └── cap.toptee.val.json
    ├── image_splits
        ├── split.dress.val.json
        ├── split.shirt.val.json
        └── split.toptee.val.json
    └── images ## images under this directory.

Json files are available in https://github.com/XiaoxiaoGuo/fashion-iq. Images are downloaded from https://github.com/postBG/CosMo.pytorch.