Project-Splinter / human_det

A Single-Human Detector runs at 70 FPS on GV100

Human Det

A Single-Human Detector runs at 70 FPS on GV100.

Install

# via pip
pip install git+https://github.com/Project-Splinter/human_det --upgrade

# via git clone
git clone https://github.com/Project-Splinter/human_det
cd human_det
python setup.py develop

Note to run demo.py, you also need to install streamer_pytorch through:

pip install git+https://github.com/Project-Splinter/streamer_pytorch --upgrade

Usage

# images
python demo.py --images <IMAGE_PATH> <IMAGE_PATH> <IMAGE_PATH> --loop --vis
# image folder
python demo.py --image_folder <IMAGE_FOLDER_PATH> --loop --vis
# videos
python demo.py --videos <VIDEO_PATH> <VIDEO_PATH> <VIDEO_PATH> --vis
# capture device
python demo.py --camera --vis

API

det_engine = Detection(device="cuda:0", fp16=False)
tensor = det_engine.load_images(image_files)
topk_bboxes, topk_probs = det_engine(tensor, class_names=["person"], topk=1)

Note: Detection is not an instance of nn.Module, so it won't be trained and updated at all.