Image classification and object detection datasets containing 20,580 images consisting of 120 breeds of dogs (Warning: Images in this dataset were taken from ImageNet).
Original dataset: images (794MB), annotations (22MB)
Image classification dataset (775MB)
Conversion from original data:
Download the images and annotations archives and extract them, resulting in the Annotation and Images directories.
Create voc directory at the same level as extracted archives.
-
Create virtual environment:
python3 -m venv venv
./venv/bin/pip install defusedxml
Run this Python script in the top-level directory to fix/move the annotations into the voc directory: ./venv/bin/python fix_dogs.py
Move all images into voc directory: mv Images/**/* voc/
Copy the images from this archive into the voc directory, replacing the originals. These are the images that we found with binary issues, that were hand edited to fix.
Run: wai-annotations convert from-voc-od -i "voc/*.xml" od-to-ic -m single to-subdir-ic -o subdir (tested with wai.annotations==0.7.5)
From inside the resulting subdir directory, run xargs -I{} rm "{}" < path/to/remove.txt with the remove.txt file to remove other damaged images (these could not be hand-fixed).
The resulting voc and subdir directories are the object detection and image classification versions of this dataset.
License (according to ImageNet website)
No, ImageNet does not own the copyright of the images. ImageNet only compiles an accurate list of web images for each synset of WordNet. For researchers and educators who wish to use the images for non-commercial research and/or educational purposes, we can provide access through our site under certain conditions and terms.