PAA-Plant-Dataset

English | 中文

PAA-Plant-Dataset is a multimodal plant dataset for plant phenotyping, multi-view 3D reconstruction, point-cloud processing, and foreground segmentation. The current release contains synchronized folder-level data for six crop types: broccoli, cabbage, cotton, rice, sunflower, and tobacco.

The release is organized around 41 plant/acquisition samples. Each sample contains RGB views, a COLMAP-compatible sparse reconstruction package, a HyperScan point cloud, a segmented HyperScan point cloud, and RGB segmentation products.

Dataset snapshot

Property	Value
Crop types	6
Plant/acquisition samples	41
Data modalities per sample	5
Data files	21,145
Total size	27.868 GB (25.954 GiB)
RGB views	4,182
HyperScan point clouds	41
Segmented HyperScan point clouds	41

Crop coverage

Published directory	Crop	Samples	Sample identifiers	Files	Size (GB)
Broccoli	Broccoli	8	Four dates × plant IDs 5, 10	4,128	5.959
Cabbage	Cabbage	8	Four dates × plant IDs 9, 21	4,126	5.272
Cotton	Cotton	5	1–5	2,578	3.027
Rice	Rice	5	1–5	2,575	2.553
Sunflower	Sunflower	7	1, 2, 3, 4, 7, 8, 9	3,612	5.472
Tobacco	Tobacco	8	Four dates × plant IDs 23, 26	4,126	5.585
Total		41		21,145	27.868

The dated sample directories use the pattern <YYYYMMDD>-<plant_id>. Examples include Broccoli/20260327-10 and Tobacco/20260421-26.

Modalities

Directory	Files	Size (GB)	Contents
Camera	4,335	4.233	Multi-view images and a COLMAP sparse model
HyperScan	41	3.138	Per-sample point cloud in TXT or PLY format
HyperScan_seg	41	1.325	Segmented per-sample point cloud in PLY format
RGB	4,182	1.757	102 numbered RGB JPEG views per sample
RGB_seg	12,546	17.415	Masks, RGBA cutouts, and white-background views
Total	21,145	27.868

Normal-map data are not included in this release.

Directory structure

PAA-Plant-Dataset/
├── Broccoli/
│   ├── 20260327-5/
│   ├── 20260327-10/
│   └── ...
├── Cabbage/
├── Cotton/
├── Rice/
├── Sunflower/
└── Tobacco/

Each sample has the following structure:

<crop>/<sample>/
├── Camera/
│   ├── images/
│   │   ├── 1.jpg
│   │   ├── 2.jpg
│   │   └── ...
│   ├── sparse/0/
│   │   ├── cameras.bin
│   │   ├── images.bin
│   │   └── points3D.bin
│   └── <sample_id>.txt          # layout note; absent in some samples
├── HyperScan/
│   └── <sample_id>.txt|ply
├── HyperScan_seg/
│   └── <sample_id>.ply
├── RGB/
│   ├── 1.jpg
│   ├── 2.jpg
│   └── ...                      # 102 views
└── RGB_seg/
    ├── masks/                   # 102 PNG masks
    ├── rgba/                    # 102 PNG foreground cutouts
    └── whitebg/                 # 102 JPEG white-background views

For tobacco samples, point-cloud and source identifiers may retain the CB- prefix, such as CB-23.ply, while the sample directory ends in -23.

File formats

Camera

Camera/images contains the image set associated with a COLMAP sparse model. The model is stored in COLMAP binary form:

• cameras.bin
• images.bin
• points3D.bin

The root TXT file in Camera, when present, is a short directory-layout note rather than a point-cloud or calibration file.

HyperScan

HyperScan contains one point cloud for each sample. The published format varies by sample:

• 17 point clouds are space-delimited TXT files.
• 24 point clouds are binary little-endian PLY files.

TXT rows contain nine values in this order:

x y z red green blue nx ny nz

The PLY headers expose the equivalent fields: XYZ coordinates, RGB color, and XYZ normals.

HyperScan_seg contains the corresponding segmented point cloud in binary PLY format for every sample.

RGB and RGB segmentation

RGB contains 102 numbered JPEG views per sample.

RGB_seg provides three aligned products for every RGB view:

• masks/<view>.png
• rgba/<view>.png
• whitebg/<view>.jpg

The repository does not claim that every segmentation mask was manually verified. Treat these masks as provided segmentation products and validate their suitability for the intended experiment.

Download

Download the full repository with the Hugging Face CLI:

hf download JackTrum938/PAA-Plant-Dataset \
  --type dataset \
  --local-dir ./PAA-Plant-Dataset

Download a subset with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="JackTrum938/PAA-Plant-Dataset",
    repo_type="dataset",
    local_dir="./PAA-Plant-Dataset",
    allow_patterns=["Broccoli/**"],
)

This is a raw folder-based dataset rather than a tabular dataset with a datasets.load_dataset() schema.

Minimal loading examples

Load a TXT HyperScan point cloud:

from pathlib import Path
import numpy as np

path = Path("PAA-Plant-Dataset/Broccoli/20260327-10/HyperScan/10.txt")
points = np.loadtxt(path)

xyz = points[:, 0:3]
rgb = points[:, 3:6].astype(np.uint8)
normals = points[:, 6:9]

Load a PLY point cloud with Open3D:

from pathlib import Path
import open3d as o3d

path = Path(
    "PAA-Plant-Dataset/"
    "Broccoli/20260416-10/HyperScan/10.ply"
)
point_cloud = o3d.io.read_point_cloud(str(path))

Load an RGB view and its mask:

from pathlib import Path
from PIL import Image

sample = Path("PAA-Plant-Dataset/Broccoli/20260327-10")
rgb = Image.open(sample / "RGB/1.jpg")
mask = Image.open(sample / "RGB_seg/masks/1.png")

Intended uses

Potential research uses include:

• Multi-view plant reconstruction
• Plant point-cloud segmentation and completion
• RGB foreground segmentation
• Plant phenotyping feature extraction
• Cross-modal RGB/point-cloud studies
• Evaluation of reconstruction robustness across crop types and growth stages

Users should define their own train, validation, and test partitions. For longitudinal crops, split by plant identity rather than individual date folders when leakage across acquisition dates would invalidate the evaluation.

License

The dataset is released under the Apache Licenses 2.0. See the license metadata on the Hugging Face repository for details.

Contact

Questions, corrections, and reports of problematic files can be submitted through the Hugging Face repository's Community tab:

https://huggingface.co/datasets/JackTrum938/PAA-Plant-Dataset/discussions