Supported Deep Object Detectors (Experimental)¶
LiGuard
allows users to add various pre-trained LIDAR and/or RGB deep object detectors into their data processing pipelines. Each deep detector usually has its dependencies. It is upto the users to install the dependencies of their cusotm deep learning algorithms. In the following, a list of currently supported (out-of-box) deep object detectors is given along their corresponding installation instructions.
Image Object Detectors¶
YOLOv5:¶
pip install -U torch==1.13.1+cu117 torchvision==0.14.1+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
pip install -U ultralytics==8.2.78
Once installation is complete, add the following entry under proc/camera/
in base_config.yml
file inside your pipeline directory.
UltralyticsYOLOv5:
enabled: True
priority: 1 # priority of process - lower is higher
model: 'yolov5s' # can be yolov5s, yolov5m, yolov5l, yolov5x https://pytorch.org/hub/ultralytics_yolov5
class_colors: # the classes you need to detect and their corresponding bbox colors in RGB format
Person: [1, 0, 0]
Bicycle: [0, 1, 0]
Car: [0, 0, 1]
Motorcycle: [0, 1, 0]
Bus: [0, 1, 1]
Truck: [1, 0, 1]
score_threshold: 0.5 # minimum score threshold for detection
Point Cloud Object Detectors¶
PointPillars:¶
pip install -U torch==1.13.1+cu117 torchvision==0.14.1+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
git clone https://github.com/zhulf0804/PointPillars /path/to/liguard/algo/nn/PointPillars
pip install -U -r /path/to/liguard/algo/nn/PointPillars/requirements.txt
cd /path/to/liguard/algo/nn/PointPillars/ops && python setup.py develop
Once installation is complete, add the following entry under proc/lidar/
in base_config.yml
file inside your pipeline directory.
PointPillarDetection:
enabled: True
priority: 1 # priority of process - lower is higher
github_repo_dir: 'algo/nn/PointPillars' # clone https://github.com/zhulf0804/PointPillars to this path and install the requirements
ckpt_file: 'algo/nn/PointPillars/pretrained/epoch_160.pth' # path to checkpoint file
score_threshold: 0.5 # minimum score threshold for detection
# Pedestrians are colored Red, Cars are colored Green, Cyclists are colored Blue