Head Detector

This is a ready-to-use head detector tool adopted yolov3.

Info

The head detector is trained on dataset HollywoodHeads, to download the original dataset, click here.

The model weights can be found in the public NAS of VIPL at 10.29.0.195:/affect/D/Data/tool/model_weights/head_detector_best.pt, or can be downloaded here.

Usage

1. Run by command line

python detect.py --model_weights path/to/model_weights.pt --input_img_folder path/to/imgs --txt_file path/to/save.txt

2. Adopted for running folders containing subfolders

model_weights = 'path/to/model_weights.pt'
model, device, imgsz = load_model(model_weights)

base_folder = 'path/to/folders'

for folder in os.listdir(base_folder):
    input_img_folder = os.path.join(base_folder, folder)
    txt_file = 'path/to/save.txt'
    dataset = LoadImages(input_img_folder, img_size=imgsz, stride=model.stride, auto=model.pt and not model.jit)
    run_single_folder(model, device, dataset, txt_file)
    LOGGER.info('Done: %s'%input_img_folder)

Output File Format

• ❌ No column header, entries are organized as ['frameID', 'label', 'xmin', 'ymin', 'xmax', 'ymax']
• ❌ No index column
• Entries of 'xmin', 'ymin', 'xmax', 'ymax' are all in 0-1 scale