abhinand4as/oxford-iiit-pet-detection

Oxford-IIIT Pet Dataset — Object Detection

Derived from the Oxford-IIIT Pet Dataset (Parkhi et al., CVPR 2012).

Task

37-class fine-grained pet head detection — tight head bounding boxes for 12 cat breeds and 25 dog breeds.

Structure

oxford_iiit_pet_dataset/
├── images/
│   ├── train/      2942 images
│   ├── val/         729 images
│   └── test/       3669 images (no bbox annotations)
├── labels/                        ← YOLO labels (Ultralytics convention)
│   ├── train/  (one .txt per image)
│   └── val/    (one .txt per image)
├── annotations/
│   ├── coco/
│   │   ├── instances_train.json
│   │   └── instances_val.json
│   └── yolo/
│       └── data.yaml
└── README.md

Why are YOLO labels in labels/ and not inside annotations/yolo/? Ultralytics automatically resolves label files by replacing images with labels in the image path. With data.yaml setting train: images/train, it looks for labels at labels/train/<stem>.txt — the top-level labels/ directory satisfies this convention directly, avoiding a redundant copy inside annotations/yolo/.

Splits

Split	Images	Annotations	Notes
train	2 942	2 943 bboxes	80 % of annotated images, stratified by breed
val	729	729 bboxes	20 % of annotated images, stratified by breed
test	3 669	—	Official test split; head bbox annotations not released

Seed: random.seed(42).

Only images with a corresponding PASCAL VOC XML head-annotation are included in train/val (~3 671 of 7 349 total images — all from the original trainval.txt).

Classes

37 breeds (class id 0-indexed, matching YOLO and COCO category_id):

ID	Breed	Species
0	Abyssinian	cat
1	Bengal	cat
2	Birman	cat
3	Bombay	cat
4	British_Shorthair	cat
5	Egyptian_Mau	cat
6	Maine_Coon	cat
7	Persian	cat
8	Ragdoll	cat
9	Russian_Blue	cat
10	Siamese	cat
11	Sphynx	cat
12	american_bulldog	dog
13	american_pit_bull_terrier	dog
14	basset_hound	dog
15	beagle	dog
16	boxer	dog
17	chihuahua	dog
18	english_cocker_spaniel	dog
19	english_setter	dog
20	german_shorthaired	dog
21	great_pyrenees	dog
22	havanese	dog
23	japanese_chin	dog
24	keeshond	dog
25	leonberger	dog
26	miniature_pinscher	dog
27	newfoundland	dog
28	pomeranian	dog
29	pug	dog
30	saint_bernard	dog
31	samoyed	dog
32	scottish_terrier	dog
33	shiba_inu	dog
34	staffordshire_bull_terrier	dog
35	wheaten_terrier	dog
36	yorkshire_terrier	dog

supercategory in COCO JSON is "cat" for ids 0–11, "dog" for ids 12–36.

Annotation formats

COCO (annotations/coco/instances_{split}.json)

Standard COCO detection format. bbox is [x, y, width, height] in pixels (top-left origin). category_id is 0-indexed (0–36).

# Detectron2
from detectron2.data.datasets import register_coco_instances
register_coco_instances("pets_train", {}, "annotations/coco/instances_train.json", "images/train")
register_coco_instances("pets_val",   {}, "annotations/coco/instances_val.json",   "images/val")

# RF-DETR / torchvision
import torchvision.datasets as tvd
ds = tvd.CocoDetection("images/train", "annotations/coco/instances_train.json")

Note: If a loader expects 1-indexed category IDs (e.g. plain Detectron2 without remapping), add 1 to category_id in the JSON or use the loader's thing_dataset_id_to_contiguous_id map.

YOLO (labels/)

One .txt per image, same stem as the image filename. Each line:

<class_id> <cx> <cy> <w> <h>

All values normalised to [0, 1]. Class IDs are 0-indexed (0–36).

data.yaml — plug-and-play for Ultralytics YOLO:

yolo train data=annotations/yolo/data.yaml model=yolo11n.pt epochs=100

Source & Citation

@InProceedings{parkhi12a,
  author       = {Omkar M. Parkhi and Andrea Vedaldi and Andrew Zisserman and C. V. Jawahar},
  title        = {Cats and Dogs},
  booktitle    = {IEEE Conference on Computer Vision and Pattern Recognition},
  year         = {2012},
}

Dataset available at https://www.robots.ox.ac.uk/~vgg/data/pets/.
For research use only — see original dataset terms of use.