Update README.md

README.md

@@ -46,7 +46,7 @@ dataset_summary: '
 
   # Note: other available arguments include ''max_samples'', etc
 
-  dataset = load_from_hub("harpreetsahota/graspclutter6d")
+  dataset = load_from_hub("Voxel51/graspclutter6d")
 
 
   # Launch the App
@@ -84,7 +84,7 @@ from fiftyone.utils.huggingface import load_from_hub
 
 # Load the dataset
 # Note: other available arguments include 'max_samples', etc
-dataset = load_from_hub("harpreetsahota/graspclutter6d")
+dataset = load_from_hub("Voxel51/graspclutter6d")
 
 # Launch the App
 session = fo.launch_app(dataset)
@@ -95,130 +95,279 @@ session = fo.launch_app(dataset)
 
 ### Dataset Description
 
-<!-- Provide a longer summary of what this dataset is. -->
+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)
+- 1 "hero" scene (scene_000109) with 13 viewpoints = 52 images for detailed exploration
+- 200 object models (.ply meshes)
+- Grasp annotations for hero scene objects only (~14 objects)
 
-- **Curated by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Language(s) (NLP):** en
-- **License:** [More Information Needed]
+- **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 [optional]
+### Dataset Sources 
 
-<!-- Provide the basic links for the dataset. -->
-
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+- **Repository:** https://huggingface.co/datasets/GraspClutter6D/GraspClutter6D
+- **Paper:** https://arxiv.org/abs/2504.06866
+- **Project Website:** https://sites.google.com/view/graspclutter6d
 
 ## Uses
 
-<!-- Address questions around how the dataset is intended to be used. -->
-
 ### Direct Use
 
-<!-- This section describes suitable use cases for the dataset. -->
-
-[More Information Needed]
+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 section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
-
-[More Information Needed]
+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
 
-<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
-
-[More Information Needed]
+### 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
 
-<!-- Motivation for the creation of this dataset. -->
+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).
 
-[More Information Needed]
+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
 
-<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
-
 #### Data Collection and Processing
 
-<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
+**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
 
-[More Information Needed]
+**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
 
-#### Who are the source data producers?
+**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
 
-<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->
+#### Who are the source data producers?
 
-[More Information Needed]
+Research teams from:
+- Korea Institute of Machinery & Materials (KIMM)
+- Gwangju Institute of Science and Technology (GIST)  
+- Korea Advanced Institute of Science and Technology (KAIST)
 
-### Annotations [optional]
+Data collected through robotic capture systems and annotated using a combination of automated methods and crowd-sourcing.
 
-<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
+### Annotations 
 
 #### Annotation process
 
-<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
+**6D Object Poses:**
+- Computed using multi-view bundle adjustment
+- Refined through iterative closest point (ICP) alignment
+- Validated against depth data
 
-[More Information Needed]
+**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.)
 
-#### Who are the annotators?
+**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
 
-<!-- This section describes the people or systems who created the annotations. -->
+#### Who are the annotators?
 
-[More Information Needed]
+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
 
-<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
-
-[More Information Needed]
+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
 
-<!-- This section is meant to convey both technical and sociotechnical 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
 
-[More Information Needed]
+**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
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+- **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
 
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
-
-## Citation [optional]
-
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
+## Citation 
 
 **BibTeX:**
 
-[More Information Needed]
+```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:**
 
-[More Information Needed]
-
-## Glossary [optional]
+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.
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
 
-[More Information Needed]
+## More Information 
 
-## More Information [optional]
+**Project Website**: https://sites.google.com/view/graspclutter6d
 
-[More Information Needed]
+**Full Dataset**: https://huggingface.co/datasets/GraspClutter6D/GraspClutter6D
 
-## Dataset Card Authors [optional]
+**Paper**: https://arxiv.org/abs/2504.06866
 
-[More Information Needed]
+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
 
-[More Information Needed]
+For questions about the full GraspClutter6D dataset: kyoobinlee@gist.ac.kr