deepannotate-ai/vegetable-harvesting-multimodal-v1

Vegetable Harvesting Multimodal Dataset v1

Overview

The Vegetable Harvesting Multimodal Dataset v1 is a publication-quality first-person (egocentric) dataset capturing hands, objects, poses, and activity annotations for agricultural manual tasks. It contains two major subset activities:

1. Vegetable Harvesting: Manual crop collection and tool usage sequence.
2. Vegetable Plucking: Fine-grained picking and placement of products.

This dataset was generated using the Hand Egocentric Multimodal Annotation Pipeline to enable research in agricultural robotics, human-hand mechanics for manual labor, Embodied AI, and temporal activity understanding.

---

Quick Facts

• Number of Activities: 2 distinct agricultural manual tasks
• Number of Videos: 6 video streams in total (3 per activity: original RGB, monocular depth, skeleton overlay)
• Total Frames: 200 processed annotations (100 per activity)
• Total Duration: 1,365.80 seconds (~22.76 minutes of master video source)
• Total Dataset Size: ~50.51 MB
• Annotation Modalities: 2D/3D hand landmarks, object bounding boxes, hand-object contacts (HOI), action classes, natural language semantic activity streams, pseudo IMU signals, trajectories, and motion statistics.

---

Included Activities

1. Vegetable_Harvesting

Represents tool-involved crop collection, including safety gear and tool bounding boxes (gloves, wrenches, etc.).

2. Vegetable_Plucking

Captures fine-grained hand-object interaction events involving fruits/vegetables (bananas, carrots, oranges, bowls) with hand-object contact classification.

---

Repository Structure

Each activity folder is self-contained and follows the same schema, making loading and processing modular.

vegetable-harvesting-multimodal-v1/
│
├── README.md                          # Dataset Card / Documentation
├── LICENSE                            # CC BY 4.0 License text
├── VERSION                            # Release version details
├── .gitattributes                     # Git LFS attributes tracking mp4, h5, and jpg files
│
├── Vegetable_Harvesting/
│     ├── rgb.mp4                      # Original harvesting video sequence
│     ├── depth.mp4                    # Estimated monocular depth video
│     ├── visualization_skeleton.mp4   # Overlay of estimated hand skeletons
│     ├── metadata.json                # Video and configuration parameters
│     ├── manifest.json                # Frame-wise indices mapping timestamps
│     ├── hand_keypoints_2d.json       # Hand joint pixel coordinates
│     ├── hand_keypoints_3d.json       # Hand joint metric coordinates
│     ├── hand_object_interactions.json# Contacts between hands and objects
│     ├── object_annotations.json      # Bounding boxes and labels of items
│     ├── actions.json                 # Frame-wise categorical action tags
│     ├── semantic_actions.json        # Descriptive natural language statements
│     ├── motion_statistics.json       # Derived physical metrics (velocity, jerk)
│     ├── pseudo_imu.json              # Synthetic accelerometer and gyroscope signals
│     ├── trajectories.json            # Motion paths of keypoints
│     ├── task_summary.json            # High-level summary of detections
│     ├── depth_metadata.json          # Monocular depth configurations
│     ├── combined_dataset.csv         # Flattened tabular csv export
│     ├── dataset.h5                   # Hierarchical binary dataset bundle
│     ├── summary_report.txt           # Text summary of logging report
│     ├── task_summary.txt             # Text summary of labels
│     └── rgb_frames/                  # 100 extracted frames as JPEG files (000001.jpg - 000100.jpg)
│
└── Vegetable_Plucking/
      ├── rgb.mp4                      # Original plucking video sequence
      ├── depth.mp4                    # Estimated monocular depth video
      ├── visualization_skeleton.mp4   # Overlay of hand skeletons
      ├── metadata.json                # Video metadata
      ├── manifest.json                # Time index mapping
      ├── hand_keypoints_2d.json       # Hand joint pixel coordinates
      ├── hand_keypoints_3d.json       # Metric 3D keypoints
      ├── hand_object_interactions.json# Contact events
      ├── object_annotations.json      # Object annotations and labels
      ├── actions.json                 # Class labels
      ├── semantic_actions.json        # Narrative labels
      ├── motion_statistics.json       # Calculated motion dynamics
      ├── pseudo_imu.json              # Synthetic IMU sensor values
      ├── trajectories.json            # Landmark paths
      ├── task_summary.json            # Extracted summary
      ├── depth_metadata.json          # Depth details
      ├── combined_dataset.csv         # Flattened CSV format
      ├── dataset.h5                   # Structured HDF5 dataset
      ├── summary_report.txt           # Execution summary report
      ├── task_summary.txt             # Labels text report
      └── rgb_frames/                  # 100 extracted frames as JPEG files (000001.jpg - 000100.jpg)

---

Annotation Files Reference

File	Type	Description
metadata.json	JSON	Configuration and capture characteristics of the input video
hand_keypoints_2d.json	JSON	Frame-aligned pixel coordinates of the 21 hand landmarks
hand_keypoints_3d.json	JSON	Relative metric 3D positions of hand joint points
hand_object_interactions.json	JSON	Mappings indicating hand-to-object contact statuses
object_annotations.json	JSON	Multi-class object bounding boxes and labels
actions.json	JSON	Coarse categorical action class tags
semantic_actions.json	JSON	Rich natural language action descriptions
motion_statistics.json	JSON	Dynamic movement details (velocity, acceleration, jerk)
pseudo_imu.json	JSON	Synthesized 6-DOF IMU acceleration and rotation profiles
trajectories.json	JSON	Position trajectories of hands and tracked items
combined_dataset.csv	CSV	Tabular CSV mapping all metrics per frame
dataset.h5	HDF5	Structured binary file containing the complete dataset

---

Dataset Statistics

Overall Staging Statistics

• Total Activities: 2
• Total Videos: 6
• Total Frames: 200 (100 per activity)
• Total Annotations: 200 frame steps
• Total Interactions: 2 active interactions
• Total Objects Detected: 373 total bounding boxes (171 harvesting, 202 plucking)
• Total Video Duration: 1,365.80 seconds
• Total Dataset Size: ~50.51 MB

Per-Activity Statistics

Vegetable_Harvesting

• Frames: 100
• Duration: 600.97 seconds (10.02 min)
• Objects Detected: 171 instances (Labels: adjustable wrench, toilet, bowl, backpack, skateboard, fire hydrant, person, laptop, cell phone, helmet, safety gloves)
• Interactions: 0 active interactions
• Activities Detected: 100 action frames (ground truth action labels logged in actions/semantic annotations)

Vegetable_Plucking

• Frames: 100
• Duration: 764.83 seconds (12.75 min)
• Objects Detected: 202 instances (Labels: orange, person, banana, frisbee, carrot, bowl, umbrella)
• Interactions: 2 active interactions
• Activities Detected: 4 categories (type_or_write, idle, hold, place) across 100 frames

---

Usage Examples (Python)

1. Parsing JSON Metadata

import json

with open("Vegetable_Harvesting/metadata.json", "r") as f:
    metadata = json.load(f)
print("Harvesting FPS:", metadata["video_info"]["fps"])

2. Loading CSV Tabular Annotation

import pandas as pd

df = pd.read_csv("Vegetable_Plucking/combined_dataset.csv")
print(df.head())

3. Reading HDF5 Dataset

import h5py

with h5py.File("Vegetable_Harvesting/dataset.h5", "r") as hf:
    print("Available keys in H5:", list(hf.keys()))
    hand_3d = hf["hand_keypoints_3d"][:]
    print("3D Landmarks shape:", hand_3d.shape)

---

Known Limitations

• Monocular Depth Estimation: Visual depth metrics (stored in depth.mp4 and depth_metadata.json) represent estimated depth extracted using a monocular estimation model, not ground-truth LiDAR or depth sensor hardware.
• Synthetic/Pseudo IMU: The IMU outputs in pseudo_imu.json are synthetic signals derived mathematically from video motion tracking statistics, rather than values measured directly from a physical IMU sensor attached to the hands or camera.

---

Future Roadmap

• v1.1: Add multi-modal depth alignment and calibrated cameras.
• v2.0: Incorporate multi-camera setups and hardware-integrated IMU sensors.
• Extended Tasks: Addition of pruning, sorting, and packaging activities.

---

Citation

@misc{vegetableharvestingmultimodalv1_2026,
  title={Vegetable Harvesting Multimodal Dataset v1},
  author={DeepAnnotate AI},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/deepannotateai/vegetable-harvesting-multimodal-v1}},
  note={Version 1.0.0}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Version History

• v1.0.0 (July 2026): Initial public release of structured annotations for Harvesting and Plucking activities.

Acknowledgements

This dataset was generated using the Hand Egocentric Multimodal Annotation Pipeline developed for this project.