bdanko/overhead-people-rgb

Overhead People RGB

Unified overhead RGB people detection dataset converted from local Roboflow YOLOv8 exports.

Images are stored as original encoded bytes without resizing, recompression, or preprocessing. All retained boxes use a single category:

• —category_id: 0
• —category: person
• —bbox: COCO-style [x, y, width, height] in pixel coordinates

Source labels such as Man, Woman, Person, ero, and 0 are preserved in objects.source_category. Source label object is omitted.

## Loading The Dataset

from io import BytesIO

from datasets import load_dataset from PIL import Image

repo_id = "bdanko/overhead-people-rgb"

Loads all parquet shards from Hugging Face.

dataset = loaddataset(repoid, split="train")

print(dataset) print(dataset[0].keys())

row = dataset[0]

Images are stored as original encoded bytes.

image = Image.open(BytesIO(row["image"]["bytes"])).convert("RGB")

COCO-style boxes in pixel coordinates: [x, y, width, height]

boxes = [obj["bbox"] for obj in row["objects"]] labels = [obj["category_id"] for obj in row["objects"]]

print(row["image_id"]) print(image.size) print(boxes[:3]) print(labels[:3])