oxford_iiit_pet

  • Description:

The Oxford-IIIT pet dataset is a 37 category pet image dataset with roughly 200 images for each class. The images have large variations in scale, pose and lighting. All images have an associated ground truth annotation of breed and species. Additionally, head bounding boxes are provided for the training split, allowing using this dataset for simple object detection tasks. In the test split, the bounding boxes are empty.

Split Examples
'test' 3,669
'train' 3,680
  • Feature structure:
FeaturesDict({
    'file_name': Text(shape=(), dtype=string),
    'head_bbox': BBoxFeature(shape=(4,), dtype=float32),
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=37),
    'segmentation_mask': Image(shape=(None, None, 1), dtype=uint8),
    'species': ClassLabel(shape=(), dtype=int64, num_classes=2),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
file_name Text string
head_bbox BBoxFeature (4,) float32
image Image (None, None, 3) uint8
label ClassLabel int64
segmentation_mask Image (None, None, 1) uint8
species ClassLabel int64 
@InProceedings{parkhi12a,
  author       = "Parkhi, O. M. and Vedaldi, A. and Zisserman, A. and Jawahar, C.~V.",
  title        = "Cats and Dogs",
  booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition",
  year         = "2012",
}