Skip to content

Datasets

Orion uses images from open source object-detection datasets to create a dataset of military vehicles and format it correctly for YOLO training. This page describes the various datasets used by Orion.

ImageNet

The first dataset Orion uses is ImageNet21k. The ImageNet21k dataset is available from the image-net website. You need to register and be granted access to download the images. We use the Winter 21 version since it gives the option of downloading the images for a single synset (a class) and we're only interested in images of specific classes (military vehicles).

Orion provides a search function to search ImageNet class names for a given query.

def search(
    keywords: list[str],
    dir: Path = settings.ORION_HOME_DIR / "imagenet",
):
    """
    Search image net classes matching the given keywords.

    Args:
        keywords (list[str]): List of keywords to search for.
        dir (Path, optional): directory where files will be downloaded.
            Defaults to ORION_HOME_DIR / "imagenet".
    """

Note

The search function will download the list of class names and ids in the dataset dir if they are not already present.

Orion also provides a convenience download function to download images and annotations for a specific class id.

def download(
    ids: list[str],
    dir: Path = settings.ORION_HOME_DIR / "imagenet",
):
    """
    Download ImageNet images and annotations for the given class ids.

    Args:
        ids (list[str]): the class ids to download.
        dir (Path, optional): the dataset directory.
            Defaults to settings.ORION_HOME_DIR / "imagenet".
    """

Note

The download function will only download images for classes that actually have object detection annotations (a lot of classes in the ImageNet21k dataset do not have annotations).

Info

Orion uses 378 annotated images from the ImageNet dataset, coming from the n04389033 class (tank, army tank, armored combat vehicle, armoured combat vehicle), which are all mapped to the AFV class.`

OpenImage

The second dataset Orion uses is Open Images which contains images with Tank detection labels. Images from OpenImages are downloaded and managed with fiftyone.

Info

Orion uses 1246 annotated images from the OpenImage dataset which are all mapped to the AFV class.

Russian Military annotated dataset

Another dataset Orion uses is the Russian Military vehicles annotated dataset provided by Tuomo Hiippala from Digital Geography Lab on Roboflow. It contains 1042 annotated images of russing military vehicles with 10 classes which we map to either the AFC or the APC class.

LABEL_MAPPING = {
    "bm-21": "AFV",
    "t-80": "AFV",
    "t-64": "AFV",
    "t-72": "AFV",
    "bmp-1": "AFV",
    "bmp-2": "AFV",
    "bmd-2": "AFV",
    "btr-70": "APC",
    "btr-80": "APC",
    "mt-lb": "APC",
}

Orion provides a download function to download the images and annotations from this dataset and structure the directory to be imported into a fiftyone dataset.

def download(dir: Path = settings.ORION_HOME_DIR / "roboflow"):
    """
    Downlad images and annotations from the russian military annotated dataset
    on roboflow and format them to be imported into a fo.Dataset.

    Args:
        dir (Path, optional): the dataset dir.
            Defaults to settings.ORION_HOME_DIR / "roboflow".
    """

Info

Orion uses 1042 annotated images from the Russian Military vehicles dataset which are mapped to the AFV or APC class.

Google Images

To improve our training dataset, we also scraped images of military vehicles from Google Image and annotated them by hand. This sample dataset is available for download from Orion's github repository and contains 669 images of vehicles from all four classes (AFV, APC, MEV and LAV).

Info

Orion uses 669 annotated images scraped from Google Images for all four classes.

The Search 2

The The Search_2 consist of 44 high-resolution digital color images of different complex natural scenes, with each scene (image) containing a single military vehicle that serves as a search target. This dataset is not used by Orion for training; it is used instead for evaluating the models on realistic long range automatic target recognition (ATR) samples.

Command-line

Orion provides a CLI command to download and setup a dataset of annotated military vehicles for training and development of automatic target recognition models. The prepare command will download images from the ImageNet, OpenImages, Russian military and Google Images sources and combine them into a single dataset on disk.

The prepare command takes as an option the directory where all the source images will be downloaded and where the full combined dataset will be saved (by default, ~/.cache/orion).

orion prepare --help

 Usage: orion prepare [OPTIONS]

 Prepare a dataset of annotated military vehicle images.

╭─ Options ───────────────────────────────────────────────────────────────────────────────────────────────╮
│ --dir   -d      DIRECTORY  Orion home directory. [default: ~/.cache/orion]                              │
│ --ids           TEXT       List of class ids to download. [default: n04389033]                          │
│ --help                     Show this message and exit.                                                  │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────╯