Pointcloud-Copy-Paste-Augmenter

Library for Copy-Paste Augmentations in Pointclouds.

• Unaugmented pointcloud:

• Pointcloud augmented with a car instance:

Requirements

• numpy
• pyyaml

File Structure

Organise your pointclouds and labels in the following way:

dataset
|
└───velodyne
│   │   file1.bin
│   │   file2.bin
|   |   ...
|
└───labels
    │   label1.label
    │   label2.label
    |   ...

In params.py, set DATA_PATH as your path to dataset. Augmented pointclouds will be saved in a similar directory structure, so create a new directory with empty velodyne and labels folders and set that to SAVE_PATH in params.py.

Usage

• Look through parameters in params.py and modify accordingly.
• Run extract_cloud.py to extract instances of a specified object class.
  • Example: To extract moving-cars, change LABEL_ID in params.py to 252.
  • Find label values in semantic-kitti.yaml under labels.
• This extracts usable instances into a folder named after the instance, folder is named moving-cars in this case.
• Change PLACEMENT_ID to IDs of possible ground locations for the extracted instances. This is where the instances will be pasted.
  • Example: If you want moving-cars to be pasted on road (40), parking (44), other-ground (49) and terrain (72), set PLACEMENT_ID to [40, 44, 49, 72].
• Change OBJ_PASTE_NAME in params.py to name of folder where instances were extracted and saved (moving-cars in this case).
• Run paste_object.py; augmented pointclouds are stored in SAVE_PATH/velodyne and augmented labels are stored in SAVE_PATH/labels.
• During the pasting sequence, an instance is randomly selected from the folder to paste into the pointcloud.
• Each instance is assigned the class label as well as an instance ID:
  • Example: A moving-car instance is augmented in and all its points are assigned the class label value of 252.
  • All its points are also assigned an instance ID based on the number of moving-cars present in the pointcloud.
  • If there are n moving-cars in the pointcloud, then the augmented points get an instance ID of n+1.
• Change NUMBER_INSTANCES if you want to augment more instances per pointcloud.
  • Example: If you want 2 instances of moving-cars pasted per pointcloud, set NUMBER_INSTANCES to 2.

NOTE: Only works with SemanticKITTI. Should work with other datasets as long as they contain instance labels, but you will have to change around some code for their instance labels if they are not stored in a similar fashion to SemanticKITTI.

Visualization

• All pointclouds are visualized using the semantic-kitti-api. Your directory structure will need to look like this for it to work:

dataset
└───sequences
    └───00
        └───velodyne
        │   │   file1.bin
        │   │   file2.bin
        |   |   ...
        |
        └───labels
            │   label1.label
            │   label2.label
            |   ...

• Install the dependencies for visualization:

sudo apt install python3-dev python3-pip python3-pyqt5.qtopengl
pip3 install --user -r requirements.txt

• To visualize the pointclouds, use visualize.py:

cd semantic-kitti-api
./visualize.py --sequence 00 --dataset /path/to/dataset/ --do_instances