Dataset Card for graspclutter6d

This is a FiftyOne dataset with 111 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/graspclutter6d")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

GraspClutter6D is a large-scale real-world dataset for robust perception and grasping in cluttered scenes. This FiftyOne dataset contains a curated subset of 99 scenes from the full GraspClutter6D dataset, optimized for visualization and exploration.

Full Dataset Statistics (original):

• 1,000 highly cluttered scenes with 14.1 objects/scene (average)
• 62.6% occlusion rate (percentage of instances with visibility ≤ 0.95)
• 200 unique objects captured in bins, shelves, and tables
• 52K RGB-D images from 4 cameras (RealSense D415, D435, Azure Kinect, Zivid)
• 736K 6D object poses with segmentation masks
• 9.3 billion 6-DoF grasp annotations

This Subset (demo):

• 99 scenes spanning the occlusion range (selected by visibility distribution)

• ~400 RGB-D images (most scenes have 1 viewpoint = 4 camera views)

• 200 object models (.ply meshes)

• Grasp annotations for hero scene objects only (~14 objects)

• Curated by: Seunghyeok Back, Joosoon Lee, Kangmin Kim, Heeseon Rho, Geonhyup Lee, Raeyoung Kang, Sangbeom Lee, Sangjun Noh, Youngjin Lee, Taeyeop Lee, and Kyoobin Lee

• Funded by : Korea Institute of Machinery & Materials (KIMM), Gwangju Institute of Science and Technology (GIST), Korea Advanced Institute of Science and Technology (KAIST)

• Language(s) (NLP): N/A (computer vision dataset)

• License:

Dataset Sources

• Repository: https://huggingface.co/datasets/GraspClutter6D/GraspClutter6D
• Paper: https://arxiv.org/abs/2504.06866
• Project Website: https://sites.google.com/view/graspclutter6d

Uses

Direct Use

This dataset is intended for:

• Research and development of robotic grasping systems in cluttered environments
• Training and evaluation of computer vision models for:
  • Instance segmentation in heavy occlusion
  • 6D object pose estimation
  • Grasp detection and planning
• Benchmarking perception algorithms against highly cluttered real-world scenarios
• Visual exploration of multi-camera RGB-D data with depth maps and 3D reconstructions
• Educational purposes to understand challenges in robotic perception

Out-of-Scope Use

This demo subset (99 scenes) is optimized for visualization and exploration in FiftyOne, not for training large-scale models. For model training, use the full GraspClutter6D dataset from HuggingFace.

Dataset Structure

FiftyOne Dataset Organization

This dataset is organized as a grouped dataset in FiftyOne, where each group represents one scene-viewpoint combination.

Group Structure: Each group has 5 slices:

• realsense_d415: RGB-D images from RealSense D415 camera
• realsense_d435: RGB-D images from RealSense D435 camera
• azure_kinect: RGB-D images from Azure Kinect camera
• zivid: RGB-D images from Zivid camera
• 3d: 3D scene reconstruction with object meshes

Camera Resolutions:

• RealSense D415/D435: 1920 × 1080
• Azure Kinect: 3840 × 2160
• Zivid: 1920 × 1200

Sample Fields (2D Camera Slices)

Field	Type	Description
scene_id	str	Scene identifier (e.g., "000109")
camera	str	Camera name (realsense_d415, realsense_d435, azure_kinect, zivid)
viewpoint	int	Viewpoint index (0-12, most scenes have only 0)
image_id	int	BOP dataset image ID
num_objects	int	Number of objects in scene
mean_visibility	float	Mean visibility fraction across all objects [0-1]
depth_scale_mm	float	Scale factor to convert depth values to millimeters
depth	Heatmap	Depth map with values in millimeters (16-bit, scaled)
detections	Detections	2D bounding boxes with instance masks and visibility metrics
grasp_lines	Polylines	Grasp approach directions (contact point + 50mm approach vector)

Detection Fields (per object instance):

• label: Object ID (e.g., "obj_066")
• bounding_box: Normalized [x, y, w, h] coordinates [0-1]
• mask: Instance segmentation mask (cropped to bbox)
• obj_id: Object type ID
• instance_idx: Instance index in scene
• visib_fract: Visible fraction of object [0-1]
• px_count_visib: Number of visible pixels
• px_count_valid: Number of valid pixels
• px_count_all: Total pixel count

Grasp Line Fields (per grasp, hero scene only):

• label: Object ID (e.g., "grasp_obj_066")
• points: Two points [[x1,y1], [x2,y2]] - contact point and tip (50mm along approach)
• obj_id: Object type ID
• grasp_score: Grasp quality score [0-1]

Sample Fields (3D Scene Slice)

Field	Type	Description
scene_id	str	Scene identifier
viewpoint	int	Viewpoint index
filepath	str	Path to .fo3d scene file
detections	Detections	3D axis-aligned bounding boxes

3D Detection Fields (per object):

• label: Object ID
• location: [x, y, z] box center in camera coordinates (mm)
• dimensions: [width, length, height] in mm
• rotation: [0, 0, 0] (axis-aligned bounding boxes)
• obj_id: Object type ID

Data Files

Per Scene:

• rgb/NNNNNN.png: RGB images (8-bit PNG)
• depth/NNNNNN.png: Depth maps (16-bit PNG, values × depth_scale_mm = mm)
• mask_visib/NNNNNN_IIIIII.png: Instance segmentation masks
• scene_camera.json: Camera intrinsics and extrinsics
• scene_gt.json: 6D object poses (rotation matrix + translation in mm)
• scene_gt_info.json: Bounding boxes and visibility metrics

Object Models:

• models_eval/obj_NNNNNN.ply: 3D meshes for 200 objects

Grasp Labels (hero scene only):

• grasp_label/obj_NNNNNN_labels.npz: Dense grasp annotations
  • points: (N, 3) contact points in object frame (meters)
  • offsets: (N, 300, 12, 4, 3) approach directions
  • collision: (N, 300, 12, 4) collision flags
  • scores: (N, 300, 12, 4) grasp quality scores [0-1]

Subset Selection Criteria

The 99 scenes were selected from 1,000 total scenes using a stratified sampling approach:

1. Scenes ranked by mean visibility (occlusion severity)
2. Divided into 3 buckets (high/medium/low occlusion)
3. 33 scenes selected from each bucket, evenly spaced by scene ID
4. Hero scene = most occluded scene with all 13 viewpoints retained
5. Other scenes limited to viewpoint 0 (4 camera views) to reduce size

Dataset Creation

Curation Rationale

The full GraspClutter6D dataset was created to address the gap between existing benchmark datasets and real-world cluttered grasping scenarios. Most existing datasets feature simple scenes with light occlusion (~35% in GraspNet-1B), while real warehouse and manufacturing environments contain densely packed objects with heavy occlusion (62.6% in this dataset).

This demo subset was curated to:

• Provide a manageable dataset size (~3-8 GB) for visualization and exploration
• Span the full range of occlusion levels (easy to hard)
• Include one detailed "hero" scene with multiple viewpoints for in-depth analysis
• Enable quick prototyping and algorithm development

Source Data

Data Collection and Processing

Capture Setup:

• 4 RGB-D cameras: RealSense D415, RealSense D435, Azure Kinect, Zivid
• Multiple viewpoints per scene (up to 13)
• Scenes arranged in bins, shelves, and tables with natural backgrounds
• 200 objects with varied shapes, sizes, textures, and materials

Annotations:

• 6D object poses obtained through multi-view optimization
• Instance segmentation masks from visible pixel analysis
• Grasp annotations computed using GraspNet framework with collision detection

This Subset Processing:

• Downloaded from HuggingFace using snapshot_download
• Extracted from 7z archives (~207 GB compressed)
• Stratified sampling by occlusion level
• Pruned to 99 scenes, 200 object models, hero scene grasp labels
• Final size: ~3-8 GB

Who are the source data producers?

Research teams from:

• Korea Institute of Machinery & Materials (KIMM)
• Gwangju Institute of Science and Technology (GIST)
• Korea Advanced Institute of Science and Technology (KAIST)

Data collected through robotic capture systems and annotated using a combination of automated methods and crowd-sourcing.

Annotations

Annotation process

6D Object Poses:

• Computed using multi-view bundle adjustment
• Refined through iterative closest point (ICP) alignment
• Validated against depth data

Instance Segmentation:

• Generated from depth discontinuities and object models
• Per-instance masks provided for visible regions
• Visibility metrics computed (visib_fract, px_count_visib, etc.)

Grasp Annotations:

• Generated using GraspNet antipodal grasp sampling
• ~18 million candidate grasps per object tested
• Collision detection performed in simulation
• Quality scores computed based on force closure metrics
• Only collision-free, high-quality grasps retained

This Subset:

• All annotations inherited from full dataset
• Grasp annotations only retained for hero scene objects

Who are the annotators?

Annotations were produced through:

• Automated pose estimation algorithms (multi-view optimization)
• Simulation-based grasp generation (GraspNet framework)
• Quality validation by research team

Personal and Sensitive Information

This dataset contains only images of inanimate objects (household items, tools, containers, etc.) arranged in controlled laboratory settings. No personal, sensitive, or private information is present.

Bias, Risks, and Limitations

Technical Limitations:

• Subset size: This demo contains only 99 of 1,000 scenes; not suitable for training large-scale models
• Grasp annotations: Only available for hero scene (~14 objects); limited grasp diversity in subset
• Viewpoint coverage: Most scenes have only 1 viewpoint (4 camera views); only hero scene has 13 viewpoints
• Object diversity: 200 objects may not cover all real-world object categories
• Environment: Captured in controlled laboratory settings; may not generalize to all real-world conditions
• Camera-specific: Depth quality and characteristics vary significantly between cameras

Dataset Biases:

• Objects selected for robotic grasping research (household items, tools); not representative of all object types
• Korean research lab environment and object selection
• Scenes arranged intentionally for high occlusion; may not reflect natural object arrangements

Recommendations

• For training: Use the full GraspClutter6D dataset from HuggingFace (1,000 scenes)
• For exploration: This subset is ideal for visualization, prototyping, and understanding dataset structure
• Multi-camera fusion: Leverage all 4 cameras for robust perception; single-camera methods may struggle
• Depth preprocessing: Account for camera-specific depth_scale_mm factors (1.0 for RealSense, 0.1 for Kinect/Zivid)
• Occlusion handling: Test algorithms on scenes across all occlusion levels (check mean_visibility field)
• Evaluation: When benchmarking, report performance separately for different occlusion ranges

Citation

BibTeX:

@article{back2025graspclutter6d,
  title={GraspClutter6D: A Large-scale Real-world Dataset for Robust Perception and Grasping in Cluttered Scenes},
  author={Back, Seunghyeok and Lee, Joosoon and Kim, Kangmin and Rho, Heeseon and Lee, Geonhyup and Kang, Raeyoung and Lee, Sangbeom and Noh, Sangjun and Lee, Youngjin and Lee, Taeyeop and Lee, Kyoobin},
  journal={arXiv preprint arXiv:2504.06866},
  year={2025}
}

APA:

Back, S., Lee, J., Kim, K., Rho, H., Lee, G., Kang, R., Lee, S., Noh, S., Lee, Y., Lee, T., & Lee, K. (2025). GraspClutter6D: A Large-scale Real-world Dataset for Robust Perception and Grasping in Cluttered Scenes. arXiv preprint arXiv:2504.06866.

More Information

Project Website: https://sites.google.com/view/graspclutter6d

Full Dataset: https://huggingface.co/datasets/GraspClutter6D/GraspClutter6D

Paper: https://arxiv.org/abs/2504.06866

Original dataset by: Seunghyeok Back, Joosoon Lee, Kangmin Kim, Heeseon Rho, Geonhyup Lee, Raeyoung Kang, Sangbeom Lee, Sangjun Noh, Youngjin Lee, Taeyeop Lee, and Kyoobin Lee

Dataset Card Contact

For questions about the full GraspClutter6D dataset: kyoobinlee@gist.ac.kr