This scripts in this folder reorganize ImageNet images and annotations into the Pytorch Dataset ImageFolder file structure.
Shared Interest is built on Pytorch. Pytorch has a class of functions called Datasets that handle loading data from disk. To load images and their bounding boxes, Shared Interest extends the Pytorch ImageFolder Dataset. ImageFolder requires images to be in a specific file structure. In order to use the data, the scripts in this folder reorganize the images and annotations, so the file structure is compatable with Pytorch ImageFolder.
For ImageNet downloaded from the ImageNet website use imagefolder_structure_from_imagenet.ipynb.
For ImageNet downloaded from Kaggle use imagefolder_structure_from_kaggle.ipynb.
After running one of these scripts, your imagenet directory will be formatted as follows:
imagenet
|
|---val/
| |---images/
| |---0000/
| |---ILSVRC2012_val_<imageid0>.jpeg
| |---ILSVRC2012_val_<imageid1>.jpeg
| |...
| |...
| |---0999/
| |---annotations/
| |---0000/
| |---ILSVRC2012_val_<imageid0>.xml
| |---ILSVRC2012_val_<imageid1>.xml
| |...
| |...
| |---0999/