Image Utils

crop_image

supervision.utils.image.crop_image(image: ImageType, xyxy: npt.NDArray[np.number] | list[int] | tuple[int, int, int, int]) -> ImageType

Crop image based on bounding box coordinates.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to crop.	required
xyxy	NDArray[number] | list[int] | tuple[int, int, int, int]	Bounding box coordinates in (x_min, y_min, x_max, y_max) format.	required

Returns:

Type	Description
ImageType	Cropped image matching input type.

Note

Coordinates are rounded to integers and clipped to the image bounds before slicing. This keeps NumPy and Pillow inputs aligned and avoids negative-index wrap-around on NumPy arrays.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
>>> image.shape
(1080, 1920, 3)
>>> xyxy = (400, 400, 800, 800)
>>> cropped_image = sv.crop_image(image=image, xyxy=xyxy)
>>> cropped_image.shape
(400, 400, 3)

>>> image = np.zeros((1920, 1080), dtype=np.uint8)
>>> image.shape
(1920, 1080)
>>> xyxy = (400, 400, 800, 800)
>>> cropped_image = sv.crop_image(image=image, xyxy=xyxy)
>>> cropped_image.shape
(400, 400)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def crop_image(
    image: ImageType,
    xyxy: npt.NDArray[np.number] | list[int] | tuple[int, int, int, int],
) -> ImageType:
    """
    Crop image based on bounding box coordinates.

    Args:
        image: The image to crop.
        xyxy:
            Bounding box coordinates in `(x_min, y_min, x_max, y_max)` format.

    Returns:
        Cropped image matching input
            type.

    Note:
        Coordinates are rounded to integers and clipped to the image bounds
        before slicing. This keeps NumPy and Pillow inputs aligned and avoids
        negative-index wrap-around on NumPy arrays.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
        >>> image.shape
        (1080, 1920, 3)
        >>> xyxy = (400, 400, 800, 800)
        >>> cropped_image = sv.crop_image(image=image, xyxy=xyxy)
        >>> cropped_image.shape
        (400, 400, 3)

        ```

        ```pycon
        >>> image = np.zeros((1920, 1080), dtype=np.uint8)
        >>> image.shape
        (1920, 1080)
        >>> xyxy = (400, 400, 800, 800)
        >>> cropped_image = sv.crop_image(image=image, xyxy=xyxy)
        >>> cropped_image.shape
        (400, 400)

        ```

    ![crop-image](https://media.roboflow.com/supervision-docs/supervision-docs-crop-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    xyxy_arr = np.asarray(xyxy, dtype=np.float64).round().astype(np.int32)
    x_min, y_min, x_max, y_max = xyxy_arr.flatten()

    if isinstance(image, np.ndarray):
        height, width = image.shape[:2]
        x_min = int(np.clip(x_min, 0, width))
        y_min = int(np.clip(y_min, 0, height))
        x_max = int(np.clip(x_max, 0, width))
        y_max = int(np.clip(y_max, 0, height))
        return image[y_min:y_max, x_min:x_max]

    if isinstance(image, Image.Image):
        width, height = image.size
        x_min = int(np.clip(x_min, 0, width))
        y_min = int(np.clip(y_min, 0, height))
        x_max = int(np.clip(x_max, 0, width))
        y_max = int(np.clip(y_max, 0, height))
        return image.crop((float(x_min), float(y_min), float(x_max), float(y_max)))

    raise TypeError(
        f"`image` must be a numpy.ndarray or PIL.Image.Image. Received {type(image)}"
    )

load_image_from_url

supervision.utils.image.load_image_from_url(value: str, cv_imread_flags: int = cv2.IMREAD_COLOR, timeout: float = 30.0, use_cache: bool = True, cache_dir: str | Path | None = None, force_reload: bool = False) -> npt.NDArray[np.uint8]

Load an image from a URL as an OpenCV image.

Parameters:

Name	Type	Description	Default
value	str	HTTP(S) URL of the image.	required
cv_imread_flags	int	OpenCV image read flag passed to cv2.imdecode. Defaults to cv2.IMREAD_COLOR.	IMREAD_COLOR
timeout	float	Request timeout in seconds. Defaults to 30.0.	30.0
use_cache	bool	If True, cache downloaded image bytes locally and reuse them on repeated calls. Defaults to True.	True
cache_dir	str | Path | None	Directory where downloaded image bytes are cached. If None, uses the system temporary directory. Defaults to None.	None
force_reload	bool	If True, re-download the image and refresh the cache. Defaults to False.	False

Returns:

Type	Description
NDArray[uint8]	Image as a NumPy array in the format selected by cv_imread_flags.

Raises:

Type	Description
ValueError	If the URL is invalid or the downloaded bytes cannot be decoded.
RequestException	If the request fails or returns an error status.

Examples:

import supervision as sv

image = sv.load_image_from_url(
    "https://media.roboflow.com/quickstart/dog.jpeg"
)
image.shape

Source code in src/supervision/utils/image.py

def load_image_from_url(
    value: str,
    cv_imread_flags: int = cv2.IMREAD_COLOR,
    timeout: float = 30.0,
    use_cache: bool = True,
    cache_dir: str | Path | None = None,
    force_reload: bool = False,
) -> npt.NDArray[np.uint8]:
    """
    Load an image from a URL as an OpenCV image.

    Args:
        value: HTTP(S) URL of the image.
        cv_imread_flags: OpenCV image read flag passed to `cv2.imdecode`.
            Defaults to `cv2.IMREAD_COLOR`.
        timeout: Request timeout in seconds. Defaults to `30.0`.
        use_cache: If `True`, cache downloaded image bytes locally and reuse them
            on repeated calls. Defaults to `True`.
        cache_dir: Directory where downloaded image bytes are cached. If `None`,
            uses the system temporary directory. Defaults to `None`.
        force_reload: If `True`, re-download the image and refresh the cache.
            Defaults to `False`.

    Returns:
        Image as a NumPy array in the format selected by `cv_imread_flags`.

    Raises:
        ValueError: If the URL is invalid or the downloaded bytes cannot be decoded.
        requests.RequestException: If the request fails or returns an error status.

    Examples:
        ```python
        import supervision as sv

        image = sv.load_image_from_url(
            "https://media.roboflow.com/quickstart/dog.jpeg"
        )
        image.shape

        ```
    """
    prepared_url = _normalize_http_url(url=value)
    if not use_cache:
        with tempfile.TemporaryDirectory() as temp_dir:
            temp_target = Path(temp_dir) / "image"
            _download_to_file(prepared_url, temp_target, timeout=timeout)
            return _decode_image_from_bytes(
                value=temp_target.read_bytes(),
                cv_imread_flags=cv_imread_flags,
            )

    cache_path = _get_image_url_cache_path(
        value=prepared_url,
        cache_dir=cache_dir,
    )
    if cache_path.exists() and not force_reload:
        try:
            return _decode_image_from_bytes(
                value=cache_path.read_bytes(),
                cv_imread_flags=cv_imread_flags,
            )
        except ValueError:
            cache_path.unlink(missing_ok=True)

    _download_to_file(prepared_url, cache_path, timeout=timeout)
    try:
        return _decode_image_from_bytes(
            value=cache_path.read_bytes(),
            cv_imread_flags=cv_imread_flags,
        )
    except ValueError:
        cache_path.unlink(missing_ok=True)
        raise

scale_image

supervision.utils.image.scale_image(image: ImageType, scale_factor: float) -> ImageType

Scale image by given factor. Scale factor > 1.0 zooms in, < 1.0 zooms out.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to scale.	required
scale_factor	float	Factor by which to scale the image.	required

Returns:

Type	Description
ImageType	Scaled image matching input type.

Raises:

Type	Description
TypeError	If image is not a numpy.ndarray or PIL.Image.Image.
ValueError	If scale factor is non-positive.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
>>> image.shape
(1080, 1920, 3)
>>> scaled_image = sv.scale_image(image=image, scale_factor=0.5)
>>> scaled_image.shape
(540, 960, 3)

>>> image = np.zeros((1920, 1080), dtype=np.uint8)
>>> image.shape
(1920, 1080)
>>> scaled_image = sv.scale_image(image=image, scale_factor=0.5)
>>> scaled_image.shape
(960, 540)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def scale_image(image: ImageType, scale_factor: float) -> ImageType:
    """
    Scale image by given factor. Scale factor > 1.0 zooms in, < 1.0 zooms out.

    Args:
        image: The image to scale.
        scale_factor: Factor by which to scale the image.

    Returns:
        Scaled image matching input
            type.

    Raises:
        TypeError: If `image` is not a `numpy.ndarray` or `PIL.Image.Image`.
        ValueError: If scale factor is non-positive.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
        >>> image.shape
        (1080, 1920, 3)
        >>> scaled_image = sv.scale_image(image=image, scale_factor=0.5)
        >>> scaled_image.shape
        (540, 960, 3)

        ```

        ```pycon
        >>> image = np.zeros((1920, 1080), dtype=np.uint8)
        >>> image.shape
        (1920, 1080)
        >>> scaled_image = sv.scale_image(image=image, scale_factor=0.5)
        >>> scaled_image.shape
        (960, 540)

        ```

    ![scale-image](https://media.roboflow.com/supervision-docs/supervision-docs-scale-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    if scale_factor <= 0:
        raise ValueError("Scale factor must be positive.")

    width_old, height_old = image.shape[1], image.shape[0]
    width_new = int(width_old * scale_factor)
    height_new = int(height_old * scale_factor)
    return cast(
        npt.NDArray[np.uint8],
        cv2.resize(image, (width_new, height_new), interpolation=cv2.INTER_LINEAR),
    )

resize_image

supervision.utils.image.resize_image(image: ImageType, resolution_wh: tuple[int, int], keep_aspect_ratio: bool = False) -> ImageType

Resize image to specified resolution. Can optionally maintain aspect ratio.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to resize.	required
resolution_wh	tuple[int, int]	Target resolution as (width, height).	required
keep_aspect_ratio	bool	Flag to maintain original aspect ratio. Defaults to False.	False

Returns:

Type	Description
ImageType	Resized image matching input type.

Raises:

Type	Description
TypeError	If image is not a numpy.ndarray or PIL.Image.Image.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
>>> image.shape
(1080, 1920, 3)
>>> resized_image = sv.resize_image(
...     image=image, resolution_wh=(1000, 1000), keep_aspect_ratio=True
... )
>>> resized_image.shape
(562, 1000, 3)

>>> image = np.zeros((1920, 1080), dtype=np.uint8)
>>> image.shape
(1920, 1080)
>>> resized_image = sv.resize_image(
...     image=image, resolution_wh=(1000, 1000), keep_aspect_ratio=True
... )
>>> resized_image.shape
(1000, 562)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def resize_image(
    image: ImageType,
    resolution_wh: tuple[int, int],
    keep_aspect_ratio: bool = False,
) -> ImageType:
    """
    Resize image to specified resolution. Can optionally maintain aspect ratio.

    Args:
        image: The image to resize.
        resolution_wh: Target resolution as `(width, height)`.
        keep_aspect_ratio: Flag to maintain original aspect ratio.
            Defaults to `False`.

    Returns:
        Resized image matching input
            type.

    Raises:
        TypeError: If `image` is not a `numpy.ndarray` or `PIL.Image.Image`.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
        >>> image.shape
        (1080, 1920, 3)
        >>> resized_image = sv.resize_image(
        ...     image=image, resolution_wh=(1000, 1000), keep_aspect_ratio=True
        ... )
        >>> resized_image.shape
        (562, 1000, 3)

        ```

        ```pycon
        >>> image = np.zeros((1920, 1080), dtype=np.uint8)
        >>> image.shape
        (1920, 1080)
        >>> resized_image = sv.resize_image(
        ...     image=image, resolution_wh=(1000, 1000), keep_aspect_ratio=True
        ... )
        >>> resized_image.shape
        (1000, 562)

        ```

    ![resize-image](https://media.roboflow.com/supervision-docs/supervision-docs-resize-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    if keep_aspect_ratio:
        image_ratio = image.shape[1] / image.shape[0]
        target_ratio = resolution_wh[0] / resolution_wh[1]
        if image_ratio >= target_ratio:
            width_new = resolution_wh[0]
            height_new = int(resolution_wh[0] / image_ratio)
        else:
            height_new = resolution_wh[1]
            width_new = int(resolution_wh[1] * image_ratio)
    else:
        width_new, height_new = resolution_wh

    return cast(
        npt.NDArray[np.uint8],
        cv2.resize(image, (width_new, height_new), interpolation=cv2.INTER_LINEAR),
    )

letterbox_image

supervision.utils.image.letterbox_image(image: ImageType, resolution_wh: tuple[int, int], color: tuple[int, int, int] | Color = Color.BLACK) -> ImageType

Resize image and pad with color to achieve desired resolution while maintaining aspect ratio.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to resize and pad. Accepts BGR arrays of shape (H, W, 3), BGRA arrays of shape (H, W, 4), grayscale arrays of shape (H, W), or a PIL Image.	required
resolution_wh	tuple[int, int]	Target resolution as (width, height).	required
color	tuple[int, int, int] | Color	Padding color. If tuple, should be in BGR format. Defaults to Color.BLACK.	BLACK

Returns:

Type	Description
ImageType	Letterboxed image matching input type.

Raises:

Type	Description
TypeError	If image is not a numpy.ndarray or PIL.Image.Image.

Note

For BGRA inputs, the alpha channel in the padding region is set to 0 (fully transparent). Grayscale inputs receive scalar padding from color[0].

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
>>> image.shape
(1080, 1920, 3)
>>> letterboxed_image = sv.letterbox_image(
...     image=image, resolution_wh=(1000, 1000)
... )
>>> letterboxed_image.shape
(1000, 1000, 3)
>>> gray = np.zeros((4, 6), dtype=np.uint8)
>>> sv.letterbox_image(image=gray, resolution_wh=(10, 10)).shape
(10, 10)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def letterbox_image(
    image: ImageType,
    resolution_wh: tuple[int, int],
    color: tuple[int, int, int] | Color = Color.BLACK,
) -> ImageType:
    """
    Resize image and pad with color to achieve desired resolution while
    maintaining aspect ratio.

    Args:
        image: The image to resize and pad. Accepts BGR arrays of shape
            ``(H, W, 3)``, BGRA arrays of shape ``(H, W, 4)``, grayscale
            arrays of shape ``(H, W)``, or a PIL ``Image``.
        resolution_wh: Target resolution as `(width, height)`.
        color: Padding color. If tuple, should be in BGR format.
            Defaults to `Color.BLACK`.

    Returns:
        Letterboxed image matching input type.

    Raises:
        TypeError: If `image` is not a `numpy.ndarray` or `PIL.Image.Image`.

    Note:
        For BGRA inputs, the alpha channel in the padding region is set to
        0 (fully transparent). Grayscale inputs receive scalar padding
        from ``color[0]``.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
        >>> image.shape
        (1080, 1920, 3)
        >>> letterboxed_image = sv.letterbox_image(
        ...     image=image, resolution_wh=(1000, 1000)
        ... )
        >>> letterboxed_image.shape
        (1000, 1000, 3)
        >>> gray = np.zeros((4, 6), dtype=np.uint8)
        >>> sv.letterbox_image(image=gray, resolution_wh=(10, 10)).shape
        (10, 10)

        ```

    ![letterbox-image](https://media.roboflow.com/supervision-docs/supervision-docs-letterbox-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    color = unify_to_bgr(color=color)
    resized_image = resize_image(
        image=image, resolution_wh=resolution_wh, keep_aspect_ratio=True
    )
    height_new, width_new = resized_image.shape[:2]
    padding_top = (resolution_wh[1] - height_new) // 2
    padding_bottom = resolution_wh[1] - height_new - padding_top
    padding_left = (resolution_wh[0] - width_new) // 2
    padding_right = resolution_wh[0] - width_new - padding_left
    image_with_borders = cast(
        npt.NDArray[np.uint8],
        cv2.copyMakeBorder(
            resized_image,
            padding_top,
            padding_bottom,
            padding_left,
            padding_right,
            cv2.BORDER_CONSTANT,
            value=color,
        ),
    )

    return image_with_borders

tint_image

supervision.utils.image.tint_image(image: ImageType, color: Color = Color.BLACK, opacity: float = 0.5) -> ImageType

Tint image with solid color overlay at specified opacity.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to tint.	required
color	Color	Overlay tint color. Defaults to Color.BLACK.	BLACK
opacity	float	Blend ratio between overlay and image (0.0-1.0). Defaults to 0.5.	0.5

Returns:

Type	Description
ImageType	Tinted image matching input type.

Raises:

Type	Description
TypeError	If image is not a numpy.ndarray or PIL.Image.Image.
ValueError	If opacity is outside range [0.0, 1.0].

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((100, 100, 3), dtype=np.uint8)
>>> tinted_image = sv.tint_image(
...     image=image, color=sv.Color.ROBOFLOW, opacity=0.5
... )
>>> tinted_image.shape
(100, 100, 3)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def tint_image(
    image: ImageType,
    color: Color = Color.BLACK,
    opacity: float = 0.5,
) -> ImageType:
    """
    Tint image with solid color overlay at specified opacity.

    Args:
        image: The image to tint.
        color: Overlay tint color. Defaults to `Color.BLACK`.
        opacity: Blend ratio between overlay and image (0.0-1.0).
            Defaults to `0.5`.

    Returns:
        Tinted image matching input
            type.

    Raises:
        TypeError: If `image` is not a `numpy.ndarray` or `PIL.Image.Image`.
        ValueError: If opacity is outside range [0.0, 1.0].

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((100, 100, 3), dtype=np.uint8)
        >>> tinted_image = sv.tint_image(
        ...     image=image, color=sv.Color.ROBOFLOW, opacity=0.5
        ... )
        >>> tinted_image.shape
        (100, 100, 3)

        ```

    ![tint-image](https://media.roboflow.com/supervision-docs/supervision-docs-tint-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    if not 0.0 <= opacity <= 1.0:
        raise ValueError("opacity must be between 0.0 and 1.0")

    overlay = np.full_like(image, fill_value=color.as_bgr(), dtype=image.dtype)
    cv2.addWeighted(
        src1=overlay, alpha=opacity, src2=image, beta=1 - opacity, gamma=0, dst=image
    )
    return image

grayscale_image

supervision.utils.image.grayscale_image(image: ImageType) -> ImageType

Convert image to 3-channel grayscale. Luminance channel is broadcast to all three channels for compatibility with color-based drawing helpers.

Parameters:

Name	Type	Description	Default
image	ImageType	The image to convert to grayscale.	required

Returns:

Type	Description
ImageType	3-channel grayscale image matching input type.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.ones((100, 100, 3), dtype=np.uint8) * 128
>>> grayscale_image = sv.grayscale_image(image=image)
>>> grayscale_image.shape
(100, 100, 3)

Source code in src/supervision/utils/image.py

@ensure_cv2_image_for_standalone_function
def grayscale_image(image: ImageType) -> ImageType:
    """
    Convert image to 3-channel grayscale. Luminance channel is broadcast to
    all three channels for compatibility with color-based drawing helpers.

    Args:
        image: The image to convert to
            grayscale.

    Returns:
        3-channel grayscale image
            matching input type.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.ones((100, 100, 3), dtype=np.uint8) * 128
        >>> grayscale_image = sv.grayscale_image(image=image)
        >>> grayscale_image.shape
        (100, 100, 3)

        ```

    ![grayscale-image](https://media.roboflow.com/supervision-docs/supervision-docs-grayscale-image-2.png){ align=center width="1000" }
    """  # noqa E501 // docs
    assert isinstance(image, np.ndarray)
    grayscaled = cast(npt.NDArray[np.uint8], cv2.cvtColor(image, cv2.COLOR_BGR2GRAY))
    return cast(npt.NDArray[np.uint8], cv2.cvtColor(grayscaled, cv2.COLOR_GRAY2BGR))

get_image_resolution_wh

supervision.utils.image.get_image_resolution_wh(image: ImageType) -> tuple[int, int]

Get image width and height as a tuple (width, height) for various image formats.

Supports both numpy.ndarray images (with shape (H, W, ...)) and PIL.Image.Image inputs.

Parameters:

Name	Type	Description	Default
image	ImageType	Input image.	required

Returns:

Type	Description
tuple[int, int]	Image resolution as (width, height).

Raises:

Type	Description
ValueError	If a numpy.ndarray image has fewer than 2 dimensions.
TypeError	If image is not a supported type (numpy.ndarray or PIL.Image.Image).

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
>>> sv.get_image_resolution_wh(image)
(1920, 1080)

Source code in src/supervision/utils/image.py

def get_image_resolution_wh(image: ImageType) -> tuple[int, int]:
    """
    Get image width and height as a tuple `(width, height)` for various image formats.

    Supports both `numpy.ndarray` images (with shape `(H, W, ...)`) and
    `PIL.Image.Image` inputs.

    Args:
        image: Input image.

    Returns:
        Image resolution as `(width, height)`.

    Raises:
        ValueError: If a `numpy.ndarray` image has fewer than 2 dimensions.
        TypeError: If `image` is not a supported type (`numpy.ndarray` or
            `PIL.Image.Image`).

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> image = np.zeros((1080, 1920, 3), dtype=np.uint8)
        >>> sv.get_image_resolution_wh(image)
        (1920, 1080)

        ```
    """
    if isinstance(image, np.ndarray):
        if image.ndim < 2:
            raise ValueError(
                "NumPy image must have at least 2 dimensions (H, W, ...). "
                f"Received shape: {image.shape}"
            )
        height, width = image.shape[:2]
        return int(width), int(height)

    if isinstance(image, Image.Image):
        width, height = image.size
        return int(width), int(height)

    raise TypeError(
        "`image` must be a numpy.ndarray or PIL.Image.Image. "
        f"Received type: {type(image)}"
    )

ImageSink

supervision.utils.image.ImageSink

Save sequential images into a directory through a context manager.

ImageSink creates the target directory on entry and writes each image using save_image, incrementing the image name pattern after every save.

Source code in src/supervision/utils/image.py

class ImageSink:
    """
    Save sequential images into a directory through a context manager.

    `ImageSink` creates the target directory on entry and writes each image
    using `save_image`, incrementing the image name pattern after every save.
    """

    def __init__(
        self,
        target_dir_path: str,
        overwrite: bool = False,
        image_name_pattern: str = "image_{:05d}.png",
    ) -> None:
        """
        Initialize context manager for saving images to directory.

        Args:
            target_dir_path: Target directory path where images will be
                saved.
            overwrite: Whether to overwrite existing directory.
                Defaults to `False`.
            image_name_pattern: File name pattern for saved images.
                Defaults to `"image_{:05d}.png"`.

        Examples:
            ```pycon
            >>> import numpy as np
            >>> import supervision as sv
            >>> import tempfile
            >>> import os
            >>> with tempfile.TemporaryDirectory() as tmpdir:
            ...     image = np.zeros((100, 100, 3), dtype=np.uint8)
            ...     with sv.ImageSink(target_dir_path=tmpdir, overwrite=True) as sink:
            ...         sink.save_image(image=image)
            ...         sink.save_image(image=image)
            ...     files = sorted(os.listdir(tmpdir))
            ...     len(files)
            2

            ```
        """
        self.target_dir_path = target_dir_path
        self.overwrite = overwrite
        self.image_name_pattern = image_name_pattern
        self.image_count = 0

    def __enter__(self) -> ImageSink:
        if os.path.exists(self.target_dir_path):
            if self.overwrite:
                shutil.rmtree(self.target_dir_path)
                os.makedirs(self.target_dir_path)
        else:
            os.makedirs(self.target_dir_path)

        return self

    def save_image(
        self, image: npt.NDArray[np.uint8], image_name: str | None = None
    ) -> None:
        """
        Save image to target directory with optional custom filename.

        Args:
            image: Image to save with shape `(height, width, 3)`
                in BGR format.
            image_name: Custom filename for saved image. If
                `None`, generates name using `image_name_pattern`. Defaults to
                `None`.

        Raises:
            OSError: If `cv2.imwrite` cannot write the image to disk.
        """
        if image_name is None:
            image_name = self.image_name_pattern.format(self.image_count)

        image_path = os.path.join(self.target_dir_path, image_name)
        if not cv2.imwrite(image_path, image):
            raise OSError(f"Failed to save image to path: {image_path}")
        self.image_count += 1

    def __exit__(
        self,
        exc_type: type[BaseException] | None,
        exc_value: BaseException | None,
        exc_traceback: TracebackType | None,
    ) -> None:
        pass

Methods:

__init__(target_dir_path: str, overwrite: bool = False, image_name_pattern: str = 'image_{:05d}.png') -> None

Initialize context manager for saving images to directory.

Parameters:

Name	Type	Description	Default
target_dir_path	str	Target directory path where images will be saved.	required
overwrite	bool	Whether to overwrite existing directory. Defaults to False.	False
image_name_pattern	str	File name pattern for saved images. Defaults to "image_{:05d}.png".	'image_{:05d}.png'

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> import tempfile
>>> import os
>>> with tempfile.TemporaryDirectory() as tmpdir:
...     image = np.zeros((100, 100, 3), dtype=np.uint8)
...     with sv.ImageSink(target_dir_path=tmpdir, overwrite=True) as sink:
...         sink.save_image(image=image)
...         sink.save_image(image=image)
...     files = sorted(os.listdir(tmpdir))
...     len(files)
2

Source code in src/supervision/utils/image.py

def __init__(
    self,
    target_dir_path: str,
    overwrite: bool = False,
    image_name_pattern: str = "image_{:05d}.png",
) -> None:
    """
    Initialize context manager for saving images to directory.

    Args:
        target_dir_path: Target directory path where images will be
            saved.
        overwrite: Whether to overwrite existing directory.
            Defaults to `False`.
        image_name_pattern: File name pattern for saved images.
            Defaults to `"image_{:05d}.png"`.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> import tempfile
        >>> import os
        >>> with tempfile.TemporaryDirectory() as tmpdir:
        ...     image = np.zeros((100, 100, 3), dtype=np.uint8)
        ...     with sv.ImageSink(target_dir_path=tmpdir, overwrite=True) as sink:
        ...         sink.save_image(image=image)
        ...         sink.save_image(image=image)
        ...     files = sorted(os.listdir(tmpdir))
        ...     len(files)
        2

        ```
    """
    self.target_dir_path = target_dir_path
    self.overwrite = overwrite
    self.image_name_pattern = image_name_pattern
    self.image_count = 0

save_image(image: npt.NDArray[np.uint8], image_name: str | None = None) -> None

Save image to target directory with optional custom filename.

Parameters:

Name	Type	Description	Default
image	NDArray[uint8]	Image to save with shape (height, width, 3) in BGR format.	required
image_name	str | None	Custom filename for saved image. If None, generates name using image_name_pattern. Defaults to None.	None

Raises:

Type	Description
OSError	If cv2.imwrite cannot write the image to disk.

Source code in src/supervision/utils/image.py

def save_image(
    self, image: npt.NDArray[np.uint8], image_name: str | None = None
) -> None:
    """
    Save image to target directory with optional custom filename.

    Args:
        image: Image to save with shape `(height, width, 3)`
            in BGR format.
        image_name: Custom filename for saved image. If
            `None`, generates name using `image_name_pattern`. Defaults to
            `None`.

    Raises:
        OSError: If `cv2.imwrite` cannot write the image to disk.
    """
    if image_name is None:
        image_name = self.image_name_pattern.format(self.image_count)

    image_path = os.path.join(self.target_dir_path, image_name)
    if not cv2.imwrite(image_path, image):
        raise OSError(f"Failed to save image to path: {image_path}")
    self.image_count += 1

Comments