Overview

GraspClutter6D is a comprehensive dataset for 1) 6-DoF grasp detection, 2) 6D object pose estimation, and 3) instance segmentation in real-world high cluttered scenes. 

Key Statistics:

• 1,000 cluttered scenes across 75 environment setups (bins, shelves, tables)

• 52,000 RGB-D images from 4 cameras (Realsense D415, D435, Azure Kinect, Zivid)

• 200 unique objects with an average of 14.1 objects per scene

• 736K 6D object pose annotations and 9.3B 6-DoF grasp poses

Resources

• Dataset [Link] - data and annotations (compatible with  GraspNet-1B and BOP dataset format)

• GraspClutter6D API [Link] - Official API for loading and evaluating 6-DoF grasps

• Annotation Tools [Link] - 6D object pose annotation toolkit

• Object & Furnitures [Link] - Purchase the physical objects and furnitures used in our dataset

• Pose Estimation Tools [comming soon] - Toolkits for 6D object pose estimation

Dataset Format

After downloading and extracting the dataset, you'll find the following directory structure:

-----

GraspClutter6D/

├── split_info/                      # Dataset splits for different evaluation setups

│   ├── grasp_train_scene_ids.json   # Train scene IDs for cross-object setup (YCB-HOPE objects)

│   ├── grasp_test_scene_ids.json    # Test scene IDs for cross-object setup (non-YCB-HOPE objects)

│   ├── ycbv_train_scene_ids.json    # Train scene IDs for intra-object setup (YCB objects)

│   ├── ycbv_test_scene_ids.json     # Test scene IDs for intra-object setup (YCB objects)

│   └── obj_ids_per_scene.json       # Metadata for custom split creation

│

├── scenes/                          # Scene data organized by scene ID

│   ├── 000000/                      # Scene 000000

│   │   ├── rgb/                     # RGB images from multiple cameras (8-bit PNG files)

│   │   │   ├── 000001.png           # Camera 1: RealSense D415 (IDs: 1,5,9,...)

│   │   │   ├── 000002.png           # Camera 2: RealSense D435 (IDs: 2,6,10,...)

│   │   │   ├── ...                  # Camera 3: Azure Kinect (IDs: 3,7,11,...)

│   │   │   └── 000052.png           # Camera 4: Zivid (IDs: 4,8,12,...)

│   │   │

│   │   ├── depth/                   # Depth images (16-bit unsigned short, in mm)

│   │   │

│   │   ├── mask/                    # Amodal masks for each instance (BOP format)

│   │   │   ├── 000001_000000.png    # Binary mask for 0th instance in image 000001

│   │   │   ├── 000001_000001.png    # Binary mask for 1st instance in image 000001

│   │   │   └── ...

│   │   │

│   │   ├── visible_mask/            # Visible masks for each instance (BOP format)

│   │   │

│   │   ├── label/                   # Semantic masks (GraspNet-1B format)

│   │   │   ├── 000001.png           # Semantic mask (0: background, 1-200: object ID)

│   │   │   └── ...

│   │   │

│   │   ├── scene_camera.json        # Camera poses, intrinsic parameters, depth scale (BOP format)

│   │   ├── scene_gt.json            # 6D object poses (BOP format)

│   │   └── scene_gt_info.json       # Metadata for 6D object metric calculation

│   │

│   ├── 000001/                      # Scene 000001

│   └── ...                          # Scenes 000002 through 000999

│

├── models/                          # High-quality 3D object models in mm units (PLY format, BOP compatible)

│   ├── obj_000001.ply

│   └── ...

│

├── models_eval/                    # Simplified 3D object models for evaluation (PLY format, mm units)

│

├── models_obj/                      # High-quality 3D object models (OBJ format, mm units)

│

├── models_obj_eval/             # Simplified 3D object models for evaluation (OBJ format, mm units)

│

├── models_m/                        # High-quality 3D object models (PLY format, meter units, GraspNet-1B format)

│

├── models_obj_m/                 # 3D object models (OBJ format, meter units, GraspNet-1B compatible)

│

├── grasp_label/                     # 6-DoF grasp pose labels (GraspNet-1B format)

│   ├── obj_000001_labels.npz

│   └── ...

│

└── dex_models/                      # Pre-processed models for fast evaluation (optional but recommended)

    ├── obj_000001.pkl

    └── ...

License

All data and resources are for only non-commercial purposes, license under Creative Commons Attribution 4.0 Non Commercial License (BY-NC-SA).