DIOCblueberry

DIOCblueberry is a dense and indiscernible object counting dataset for in-field blueberry scenes. It is organized for Hugging Face as an imagefolder dataset with per-split metadata.jsonl files. Our benchmark is publicly available at https://huggingface.co/datasets/weihao-bo/DIOCblueberry.

Dataset Description

The dataset matches the paper description for DIOCblueberry:

• 6,265 high-resolution images
• 679,030 point annotations
• 108.38 annotated instances per image on average
• Dense, small, visually ambiguous blueberry targets under natural field conditions
• Point annotations plus exemplar box annotations in box_examples_coordinates

The original archive also contained a DIOC16 directory. That directory was not used for this Hugging Face dataset because it contains 1,677 mixed-category images and does not match the paper's DIOCblueberry statistics.

Repository Layout

data/
  train/
    *.jpg
    metadata.jsonl
  val/
    *.jpg
    metadata.jsonl
  test/
    *.jpg
    metadata.jsonl
annotation.json
train_val_test.json
images_class.txt

Each metadata.jsonl row contains:

• file_name: image filename relative to its split folder
• image_id: original image filename
• width: image width from the annotation file
• height: image height from the annotation file
• count: number of annotated blueberry points in the image
• points: list of [x, y] point annotations
• box_examples_coordinates: exemplar box coordinates from the original annotation file

Splits

The canonical split uses the source file train_val_test_1253.json, renamed here to train_val_test.json because it covers all 6,265 images described in the paper.

Split	Images
train	3,759
val	1,253
test	1,253
total	6,265

Loading Example

from datasets import load_dataset

dataset = load_dataset("weihao-bo/DIOCblueberry", data_dir="data")
print(dataset)
print(dataset["train"][0].keys())

Notes

The annotations are intended for dense object counting. For density-map training, convert the points field into sparse impulse maps and apply the desired Gaussian kernel or adaptive kernel policy. The original annotation.json is kept at the repository root for compatibility with code that expects the paper-style source format.