Detic image labaler#

Detic is a deep learning algorithum developed by facebook research. This tool generate annotation data by using detic. Detic can classify 21k classes. This tools are running onnx converted detic models with opset=16 in this repository. Thank you for ailia-models developers.

Use with CLI#

Warning

This sample command is written with the assumption that it will be executed in the root directory of the amber package.

amber automation detic_image_labeler tests/automation/detic_image_labeler.yaml tests/rosbag/ford/read_image.yaml tests/rosbag/ford/ford.mcap output.mcap

Task description yaml for the detic_image_labaler is here.

confidence_threshold: 0.5      # If the confidence overs the threshold, detic determines the object are exists.
video_output_path: output.mp4  # Relative path to the visualization result.
vocabulary: "lvis"             # Vocabulary of detic, you can choose from "lvis" and "imagenet_21k"
model_type: "SwinB_896_4x"     # Model type of detic, you can choose from "SwinB_896_4x" and "R50_640_4x"

After executing this command, output.mp4 movie file was generated.

Use with Python API#

current_path = Path(os.path.dirname(os.path.realpath(__file__)))
labeler = DeticImageLabeler(str(current_path / "automation" / "detic_image_labeler.yaml"))
dataset = ImagesDataset(
    str(current_path / "rosbag" / "ford" / "ford.mcap"),
    str(current_path / "rosbag" / "ford" / "read_image.yaml"),
)
labeler.inference(dataset)

detic_image_labeler.yaml and read_image.yaml are exactly same when you use detic_image_labaler with CLI.

After executing this command, output.mp4 movie file was generated.