This script is designed to simulate real world 3D scanners (Kinect, PrimeSense, etc.) in virtual environments. With this tool, you can easily create RGB-D or point cloud datasets from synthetic models, such datasets could be useful in the field related to computer graphics and computer vision.

Given a 3D mesh, a set of customizable virtual cameras scan the model from different views, the RGB-D images and corresponding point clouds are then produced. The dimension of the output point cloud is N×7, which contains xyz coordinate, rgb color, and an instance index (if available).

The script is running on Blender (tested on version 2.80, 2.90), the render core is from bpycv.

• Python environment of Blender

  Windows:

  cd <path to blender>/2.90/python/bin
  python -m ensurepip
  python -m pip install -U pip setuptools wheel
  

  Linux:

  cd <path to blender>/2.90/python/bin
  ./python3.7m -m ensurepip
  ./python3.7m -m pip install -U pip setuptools wheel
  

• bpycv and other dependencies

  Windows:

  # Download openexr for windows from https://www.lfd.uci.edu/~gohlke/pythonlibs/#openexr
  python -m pip install OpenEXR-1.3.2-cp37-cp37m-win_amd64.whl
  python -m pip install opencv-python bpycv
  

  Linux:

  ./python3.7m -m pip install opencv-python openexr bpycv
  

• h5py (optional), save point cloud compactly

  Windows:

  python -m pip install h5py
  

  Linux:

  ./python3.7m -m pip install h5py
  

Two different setups are provided: scan_single_around.py and scan_single_random.py. The former creates fixed cameras from six directions around the object, pursuing a complete coverage of the object; the latter creates random cameras in each frame, which simulate the real scanning situation.

Run the following command:

blender -b -P scan_single_xxx.py <model_dir> <output_dir> <model_id> <save_rgbd_image> <save_pc_per_view> <save_pc_complete> <pc_per_view_size> <pc_complete_size>

Example: blender -b -P scan_single_around.py ./example_obj/model.obj ./output plane 1 1 1 2048 16384

Run the following command:

python scan_batch_shapenet.py <num_process> <dataset_dir> <output_dir> <save_rgbd_image> <save_pc_per_view> <save_pc_complete> <pc_per_view_size> <pc_complete_size>

Example: python scan_batch_shapenet.py 8 F:/datasets/ShapeNetCore.v2 ./output 1 1 1 2048 16384

Note: This script is tested on ShapeNetCore.v2 dataset. For other datasets, please modify the directory structure in scan_batch_shapenet.py accordingly.

• Single scan

  • model_dir: the full path to the target model
  • model_id: the unique name of the target model

• Batch scan

  • num_process: number of processors to be used, adjust according to your computer hardware
  • dataset_dir: the full path to the dataset root folder

• Common

  • output_dir: the output directory
  • save_rgbd_image: whether to save RGB-D and instance image (0 for false, 1 for true, same for below)
  • save_pc_per_view: whether to save the partial point cloud of each single view
  • save_pc_complete: whether to save the complete point cloud by combining all views
  • pc_per_view_size: the points number of the single view point cloud
  • pc_complete_size: the points number of the complete point cloud

• Camera intrinsic:

  Modify INTRINSIC matrix in scan_single_around.py or scan_single_random.py

• Camera numbers:

  Modify cam_locations list in scan_single_around.py or NUM_FRAMES value in scan_single_random.py

• Lights:

  Modify setup_blender() function in scan_single_around.py or scan_single_random.py