dataset.md

Data Preparation

1. Download the images from HAKE dataset. You may follow the official instruction. For your convenience, you may download all the images required by Bongard-HOI here. The images should be extracted to ./assets/data/hake/images and the file structure looks like:

   data
   └── hake
       └── images
           ├── hake_images_20190730
           ├── hcvrd
           ├── hico_20160224_det
           │   └── images
           │       ├── test2015
           │       └── train2015
           ├── openimages
           │   └── images
           ├── pic
           │   └── image
           │       ├── train
           │       └── val
           └── vcoco
               ├── train2014
               └── val2014
   

2. Download the Bongard-HOI annotations from here and extract them to ./cache

3. Download the detected bounding boxes from here and extract them to ./cache

4. Download the pretrained ResNet-50 from here and extract them to ./cache