COCO evaluation sets

Object-detection datasets in the COCO label space, stored as parquet. Every row is one image with its boxes.

config	split	images
coco2017	train	118287
coco2017	validation	5000
voc2007	train	2501
voc2007	validation	2510
voc2007	test	4952
voc2012	train	5717
voc2012	validation	5823
visdrone2019	train	6471
visdrone2019	validation	548
visdrone2019	test	1610

Columns

• image: the image
• image_id, file_name, width, height: same as the COCO images entry
• coco2017: original COCO 2017 train2017 / val2017, original image_id and annotations (with segmentation)
• voc2007 / voc2012: VOC classes mapped to COCO category ids and names; boxes are [xmin-1, ymin-1, w, h]; difficult objects have iscrowd=1 so COCO eval ignores them. Splits are the official VOC train / val / test lists; VOC2012 has no test because its labels are not public
• visdrone2019: VisDrone2019-DET train / val / test (= test-dev; test-challenge labels are not public), drone view, 69% small objects. pedestrian/people -> person, motor -> motorcycle, van -> car (plus a truck ignore box), tricycle/awning-tricycle/others and ignored regions -> ignore boxes (iscrowd=1). Only 6 COCO classes exist (person, bicycle, car, motorcycle, bus, truck). Many images have more than 100 objects, so COCO's default maxDets=100 caps recall
• objects: list of annotations with id, category_id, category (name), bbox (COCO [x, y, w, h]), area, iscrowd, segmentation (original COCO segmentation as a JSON string; "[]" for VOC and VisDrone)

The original COCO json files are in annotations/ for evaluation with pycocotools.

License

This repository only re-formats the original data; all rights stay with the original authors. COCO: annotations CC BY 4.0, images under their Flickr licenses. PASCAL VOC: images are subject to the Flickr terms of use. VisDrone: academic, non-commercial research use only.

Usage

from datasets import load_dataset

ds = load_dataset("Hyeon1222/COCO", "visdrone2019", split="validation")
ex = ds[0]
ex["image"], ex["objects"]["bbox"], ex["objects"]["category_id"]

load_dataset(..., split=...) downloads every split of the config. To fetch one split only (e.g. COCO val without the 19GB train), pass its files:

ds = load_dataset("Hyeon1222/COCO", data_files={"validation": "data/coco2017/validation-*.parquet"}, split="validation")

PyTorch loader (code/dataset_loaders.py)

A single-file loader that turns every split above into a torchvision-style detection dataset (image tensor at original size, target with boxes in [x1, y1, x2, y2], labels, image_id, iscrowd, area). It downloads only the split you ask for, can stream, and gives the matching COCO ground truth for pycocotools.

pip install torch torchvision pillow datasets huggingface_hub pycocotools

import os, sys
from huggingface_hub import hf_hub_download

p = hf_hub_download("Hyeon1222/COCO", "code/dataset_loaders.py", repo_type="dataset")
sys.path.insert(0, os.path.dirname(p))
from dataset_loaders import build_dataset, make_loader, get_coco_gt, to_coco_results

ds = build_dataset("voc2007_test")                       # label_space="coco91" (COCO category_id, default) or "coco80" (1..80)
loader = make_loader(ds, batch_size=4)
for images, targets in loader:
    ...

ds = build_dataset("coco2017_train", streaming=True, max_images=1000)   # first 1000 images without downloading 19GB

# evaluation: model outputs -> COCO results -> pycocotools
coco_gt = get_coco_gt(ds)                                # annotations/*.json of that split
results = to_coco_results(outputs, targets)              # outputs: [{"boxes", "labels", "scores"}, ...]

loader name	config / split
coco2017_train, coco2017_val	coco2017 / train, validation
voc2007_train, voc2007_val, voc2007_test	voc2007 / train, validation, test
voc2012_train, voc2012_val	voc2012 / train, validation
visdrone2019_train, visdrone2019_val, visdrone2019_test	visdrone2019 / train, validation, test

Boxes with iscrowd=1 are ignore regions: they are kept in target (drop them with drop_crowd=True) and pycocotools does not count detections on them as false positives. The loader also reads local COCO-format datasets (Open Images with federated verified_labels, BDD100K, COCO-O, custom) through build_dataset(name) once they are registered in a registry.json.