A toolkit for generating synthetic datasets for 6DOF object pose estimation

• Multi-format output: Generates datasets compatible with DOPE, EfficientPose, YOLO6DPose, and BOP formats

• Background randomization: Supports both HDR and regular background images with automatic augmentation

conda env update --name 6dof_datasetgen --file 6dof_datasetgen_env.yml --prune
conda activate 6dof_datasetgen

blenderproc quickstart  # This will install everything to work with BlenderProc

blenderproc pip install pyquaternion tqdm

Download 3D models and place them in the models/ folder:

• Download link: 3D Models
• Supported formats: .ply files
• Naming convention: obj_1.ply, obj_2.ply, etc.

Download background images and place them in the backgrounds/ folder:

• HDRI source: Poly Haven HDRIs
• Supported formats: .hdr, .png, .jpg, .jpeg
• The toolkit automatically detects HDR vs regular images

Generate synthetic data with objects moving randomly around the scene:

blenderproc run syntetic_data_example0.py

Generate synthetic data with precise object control using utility functions:

blenderproc run syntetic_data_example1.py

The toolkit includes specialized functions for object manipulation (available at line 610+ in example1):

• close_up_position(): Centers objects with Y-axis variation only
• close_up_left_right_position(): Positions objects with X and Y axis control
• turn_around_rotation_during_frames(): Rotates objects 360° across all frames
• turn_around_rotation(): Applies fixed rotation angles

The toolkit generates multiple dataset formats simultaneously:

datasets/dope_data/
├── 000000.png
├── 000000.json
└── ...

datasets/YOLO6DPose_data/
├── JPEGImages/
├── labels/        # 29-value format per object
└── mask/

datasets/bop_data/lm/train_pbr/000000/
├── rgb/
├── depth/
├── mask/
├── scene_gt.json
├── scene_gt_info.json
└── scene_camera.json

datasets/efficientpose_data/data/01/
├── rgb/
├── depth/
├── mask/
├── gt.yml
├── info.yml
├── train.txt
└── test.txt

Key parameters in the main scripts:

nb_frames = 5           # Number of frames to generate
width = 512            # Image width
height = 512           # Image height
scale = 0.01           # Object scale (meters: 0.01, cm: 1.0, mm: 0.001)
target_obj_idx = 0     # Target object index for YOLO6DPose and EfficientPose

After running the generation, you'll find:

datasets/
├── dope_data/           # DOPE format
├── YOLO6DPose_data/     # YOLO6DPose format
├── bop_data/            # BOP format (intermediate)
└── efficientpose_data/  # EfficientPose format