ViViD-5k: Vineyard Vision Dataset

ViViD-5k is a large-scale vineyard image dataset for grape cluster detection, segmentation, and berry counting. Every image is annotated with instance masks and bounding boxes for grape clusters, plus a centroid keypoint for every visible berry — making it one of the densest object-counting datasets in agriculture.

📄 Paper: arXiv:2605.24353 · 💻 Code: GrapeSAM · 📓 Demo: notebooks/demo.ipynb

At a glance

Images	5,000 (train 4,000 / val 500 / test 500)
Grape varieties	13
Berry keypoints	648,710
Cluster instances (masks + bboxes)	18,561
Berries per image	130 mean / 2,144 max
Resolution	up to 4,650 px (longest side)

About 2% of images (110/5,000) contain no berry keypoints — negative or fully-occluded examples; 97 of those also have zero cluster instances.

Dataset statistics

More statistics (clusters per image, cluster sizes, spatial density)

Quick start

from huggingface_hub import hf_hub_download
import numpy as np
from PIL import Image

repo = "XZhi/ViViD-5k"
img = Image.open(hf_hub_download(repo, "data/imgs/0.png", repo_type="dataset"))
pts = np.load(hf_hub_download(repo, "data/anns/points/0.npy", repo_type="dataset"))
print(f"{len(pts)} berries annotated")

Or browse interactively: the Dataset Viewer above shows every image with its berry and cluster counts (previews are downscaled to 1024 px; full-resolution images live in data/imgs/).

Structure

data/
├── imgs/                      # 5,000 raw images (JPG, PNG), full resolution
├── anns/
│   ├── instances_updated.json # COCO-format cluster instance masks and bboxes
│   └── points/                # <img_id>.npy — (N, 2) berry centroid coordinates
├── metadata.csv               # per-image: split, berry_count, cluster_count, width, height
├── train.txt / val.txt / test.txt   # split definitions
parquet/                       # Dataset Viewer previews (1024px) + counts
assets/                        # dataset card figures
scripts/                       # reproducible generators for metadata, figures, parquet
notebooks/                     # demo notebook (Colab-ready)
docs/                          # promotion checklist

Known issues

3 source JPEGs are truncated (1597915859896.jpg, 1599030751681.jpg, 1599042714933.jpg) — decodable except for their final rows. Use from PIL import ImageFile; ImageFile.LOAD_TRUNCATED_IMAGES = True when reading raw images (Dataset Viewer previews already handle this).

Citation

If you use ViViD-5k, please cite:

@misc{tong2026vivid5kvineyardvisiondataset,
      title={ViViD-5K: Vineyard vision dataset for field-based berry detection and segmentation and grape cluster closure estimation}, 
      author={Xiangzhi Tong and Chengrui Zhang and Mac Flaherty and Andre Matteo Garcia and Dominic Gorman and Jonathan Jaramillo and Justine E. Vanden Heuvel and Yu Jiang},
      year={2026},
      eprint={2605.24353},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.24353}, 
}

License

CC-BY-4.0