Derived Segmentors

BrightnessSegmentor

Bases: Segmentor

Source code in geograypher/predictors/derived_segmentors.py

class BrightnessSegmentor(Segmentor):
    def __init__(self, brightness_threshold: float = np.sqrt(0.75)):
        self.brightness_threshold = brightness_threshold
        self.num_classes = 2

    def segment_image(self, image: np.ndarray, **kwargs):
        image_brightness = np.linalg.norm(image, axis=-1)
        thresholded_image = image_brightness > self.brightness_threshold
        class_index_image = thresholded_image.astype(np.uint8)
        one_hot_image = self.inds_to_one_hot(class_index_image)
        return one_hot_image

LookUpSegmentor

Bases: Segmentor

Source code in geograypher/predictors/derived_segmentors.py

class LookUpSegmentor(Segmentor):
    def __init__(self, base_folder, lookup_folder, num_classes=10):
        self.base_folder = Path(base_folder)
        self.lookup_folder = lookup_folder
        self.num_classes = num_classes

    def segment_image(self, image: np.ndarray, filename: PATH_TYPE, image_scale: float):
        relative_path = Path(filename).relative_to(self.base_folder)
        lookup_path = Path(self.lookup_folder, relative_path)
        lookup_path = lookup_path.with_suffix(".png")

        image = imread(lookup_path)
        if image_scale != 1:
            image = resize(
                image,
                (int(image.shape[0] * image_scale), int(image.shape[1] * image_scale)),
                order=0,  # Nearest neighbor interpolation
            )
        one_hot_image = self.inds_to_one_hot(image, num_classes=self.num_classes)
        return one_hot_image

TabularRectangleSegmentor

Bases: Segmentor

Source code in geograypher/predictors/derived_segmentors.py

class TabularRectangleSegmentor(Segmentor):
    def __init__(
        self,
        detection_file_or_folder: PATH_TYPE,
        image_shape: tuple,
        label_key: str = "instance_ID",
        image_path_key: str = "image_path",
        imin_key: str = "ymin",
        imax_key: str = "ymax",
        jmin_key: str = "xmin",
        jmax_key: str = "xmax",
        detection_file_extension: str = "csv",
        strip_image_extension: bool = False,
        use_absolute_filepaths: bool = False,
        split_bbox: bool = True,
        image_folder: typing.Union[PATH_TYPE, None] = None,
    ):
        """Lookup rectangular bounding boxes corresponding to detections from a CSV or folder of them.

        Args:
            detection_file_or_folder (PATH_TYPE):
                Path to the CSV file with detections or a folder thereof
            image_shape (tuple):
                The (height, width) shape of the image in pixels.
            label_key (str, optional):
                The column that corresponds to the class. Defaults to "label".
            image_path_key (str, optional):
                The column that has the image filename. Defaults to "image_path".
            imin_key (str, optional):
                Column of the minimum i dimension. Defaults to "ymin".
            imax_key (str, optional):
                Column of the max i dimension. Defaults to "ymax".
            jmin_key (str, optional):
                Column of the min j dimension. Defaults to "xmin".
            jmax_key (str, optional):
                Column of the max j dimension. Defaults to "xmax".
            detection_file_extension (str, optional):
                File extension of the detection files. Defaults to "csv".
            strip_image_extension (bool, optional):
                Remove the extension from the image filenames. Defaults to True.
            use_absolute_filepaths (bool, optional):
                Add the absolute path from the image folder to the filenames. Defaults to False.
            split_bbox (bool, optional):
                Split the bounding box from one column rather than having seperate columns for imin,
                imax, jmin, jmax. Defaults to True.
            image_folder (PATH_TYPE, optional): Path to the image folder. Defaults to None.
        """
        self.image_shape = image_shape

        self.label_key = label_key
        self.image_path_key = image_path_key
        self.imin_key = imin_key
        self.imax_key = imax_key
        self.jmin_key = jmin_key
        self.jmax_key = jmax_key
        self.split_bbox = split_bbox

        # Load the detections
        self.labels_df = self.load_detection_files(
            detection_file_or_folder=detection_file_or_folder,
            detection_file_extension=detection_file_extension,
            image_folder=image_folder,
            use_absolute_filepaths=use_absolute_filepaths,
            strip_image_extension=strip_image_extension,
            image_path_key=image_path_key,
        )

        # Group the predictions
        self.grouped_labels_df = self.labels_df.groupby(by=self.image_path_key)

        # List the images
        self.image_names = list(self.grouped_labels_df.groups.keys())
        # Record the class names and number of classes
        self.class_names = np.unique(self.labels_df[self.label_key]).tolist()
        self.num_classes = len(self.class_names)

    def load_detection_files(
        self,
        detection_file_or_folder: PATH_TYPE,
        detection_file_extension: str,
        image_folder: PATH_TYPE,
        use_absolute_filepaths: bool,
        strip_image_extension: bool,
        image_path_key: str,
    ):
        # Determine whether the input is a file or folder
        if Path(detection_file_or_folder).is_file():
            # If it's a file, make a one-length list
            files = [detection_file_or_folder]
        else:
            # List all the files in the folder with the requested extesion
            files = sorted(
                Path(detection_file_or_folder).glob("*" + detection_file_extension)
            )

        # Read the individual files
        dfs = [pd.read_csv(f) for f in files]

        # Concatenate the dataframes into one
        labels_df = pd.concat(dfs, ignore_index=True)

        # Add an sequential instance ID column if not present
        if "instance_ID" not in labels_df.columns:
            labels_df["instance_ID"] = labels_df.index

        # Prepend the image folder to the image filenames if requested to make an absolute filepath
        if image_folder is not None and use_absolute_filepaths:
            absolute_filepaths = [
                str(Path(image_folder, img_path))
                for img_path in labels_df[image_path_key].tolist()
            ]
            labels_df[image_path_key] = absolute_filepaths

        # Strip the extension from the image filenames if requested
        if strip_image_extension:
            image_path_without_ext = [
                str(Path(img_path).with_suffix(""))
                for img_path in labels_df[image_path_key].tolist()
            ]
            labels_df[image_path_key] = image_path_without_ext

        return labels_df

    def get_all_detections(self) -> pd.DataFrame:
        """Return the aggregated detections dataframe"""
        return self.labels_df

    def save_detection_data(self, output_csv_file: PATH_TYPE):
        """Save the aggregated detections to a file

        Args:
            output_csv_file (PATH_TYPE):
                A path to a CSV file to save the detections to. The containing folder will be
                created if needed.
        """
        ensure_containing_folder(output_csv_file)
        self.labels_df.to_csv(output_csv_file)

    def get_corners(self, data, as_int=True):
        if self.split_bbox:
            # TODO split row
            bbox = data["bbox"]
            bbox = bbox[1:-1]
            splits = bbox.split(", ")
            jmin, imin, width, height = [float(s) for s in splits]

            imax = imin + height
            jmax = jmin + width

            imin = imin
            jmin = jmin
        else:
            imin = data[self.imin_key]
            imax = data[self.imax_key]
            jmin = data[self.jmin_key]
            jmax = data[self.jmax_key]

        corners = imin, jmin, imax, jmax
        if as_int:
            corners = list(map(int, corners))

        return corners

    def segment_image(self, image, filename, image_scale, vis=False):
        output_shape = self.image_shape
        label_image = np.full(output_shape, fill_value=np.nan, dtype=float)

        name = filename.name
        if name in self.image_names:
            df = self.grouped_labels_df.get_group(name)
        # Return an all-zero segmentation image
        else:
            return label_image

        for _, row in df.iterrows():
            label = row[self.label_key]
            label_ind = self.class_names.index(label)

            imin, jmin, imax, jmax = self.get_corners(
                row,
            )

            label_image[imin:imax, jmin:jmax] = label_ind

        if vis:
            plt.imshow(label_image, vmin=0, vmax=10, cmap="tab10")
            plt.show()

        if image_scale != 1.0:
            output_size = (int(image_scale * x) for x in label_image.shape[:2])
            label_image = resize(label_image, output_size, order=0)

        return label_image

    def get_detection_centers(self, filename):
        """_summary_

        Args:
            filename (_type_): _description_

        Returns:
            _type_: (n,2) array for (i,j) centers for each detection
        """
        if filename not in self.image_names:
            # Empty array of detection centers
            return np.zeros((0, 2))

        # Extract the corresponding dataframe
        df = self.grouped_labels_df.get_group(filename)

        all_corners = []
        for _, row in df.iterrows():
            corners = self.get_corners(row, as_int=False)
            all_corners.append(corners)

        all_corners = zip(*all_corners)
        all_corners = [np.array(x) for x in all_corners]

        imin, jmin, imax, jmax = all_corners

        # Average the left-right, top-bottom pairs
        centers = np.vstack([(imin + imax) / 2, (jmin + jmax) / 2]).T
        return centers

Functions

__init__(detection_file_or_folder, image_shape, label_key='instance_ID', image_path_key='image_path', imin_key='ymin', imax_key='ymax', jmin_key='xmin', jmax_key='xmax', detection_file_extension='csv', strip_image_extension=False, use_absolute_filepaths=False, split_bbox=True, image_folder=None)

Lookup rectangular bounding boxes corresponding to detections from a CSV or folder of them.

Parameters:

Name	Type	Description	Default
detection_file_or_folder	PATH_TYPE	Path to the CSV file with detections or a folder thereof	required
image_shape	tuple	The (height, width) shape of the image in pixels.	required
label_key	str	The column that corresponds to the class. Defaults to "label".	'instance_ID'
image_path_key	str	The column that has the image filename. Defaults to "image_path".	'image_path'
imin_key	str	Column of the minimum i dimension. Defaults to "ymin".	'ymin'
imax_key	str	Column of the max i dimension. Defaults to "ymax".	'ymax'
jmin_key	str	Column of the min j dimension. Defaults to "xmin".	'xmin'
jmax_key	str	Column of the max j dimension. Defaults to "xmax".	'xmax'
detection_file_extension	str	File extension of the detection files. Defaults to "csv".	'csv'
strip_image_extension	bool	Remove the extension from the image filenames. Defaults to True.	False
use_absolute_filepaths	bool	Add the absolute path from the image folder to the filenames. Defaults to False.	False
split_bbox	bool	Split the bounding box from one column rather than having seperate columns for imin, imax, jmin, jmax. Defaults to True.	True
image_folder	PATH_TYPE	Path to the image folder. Defaults to None.	None

Source code in geograypher/predictors/derived_segmentors.py

def __init__(
    self,
    detection_file_or_folder: PATH_TYPE,
    image_shape: tuple,
    label_key: str = "instance_ID",
    image_path_key: str = "image_path",
    imin_key: str = "ymin",
    imax_key: str = "ymax",
    jmin_key: str = "xmin",
    jmax_key: str = "xmax",
    detection_file_extension: str = "csv",
    strip_image_extension: bool = False,
    use_absolute_filepaths: bool = False,
    split_bbox: bool = True,
    image_folder: typing.Union[PATH_TYPE, None] = None,
):
    """Lookup rectangular bounding boxes corresponding to detections from a CSV or folder of them.

    Args:
        detection_file_or_folder (PATH_TYPE):
            Path to the CSV file with detections or a folder thereof
        image_shape (tuple):
            The (height, width) shape of the image in pixels.
        label_key (str, optional):
            The column that corresponds to the class. Defaults to "label".
        image_path_key (str, optional):
            The column that has the image filename. Defaults to "image_path".
        imin_key (str, optional):
            Column of the minimum i dimension. Defaults to "ymin".
        imax_key (str, optional):
            Column of the max i dimension. Defaults to "ymax".
        jmin_key (str, optional):
            Column of the min j dimension. Defaults to "xmin".
        jmax_key (str, optional):
            Column of the max j dimension. Defaults to "xmax".
        detection_file_extension (str, optional):
            File extension of the detection files. Defaults to "csv".
        strip_image_extension (bool, optional):
            Remove the extension from the image filenames. Defaults to True.
        use_absolute_filepaths (bool, optional):
            Add the absolute path from the image folder to the filenames. Defaults to False.
        split_bbox (bool, optional):
            Split the bounding box from one column rather than having seperate columns for imin,
            imax, jmin, jmax. Defaults to True.
        image_folder (PATH_TYPE, optional): Path to the image folder. Defaults to None.
    """
    self.image_shape = image_shape

    self.label_key = label_key
    self.image_path_key = image_path_key
    self.imin_key = imin_key
    self.imax_key = imax_key
    self.jmin_key = jmin_key
    self.jmax_key = jmax_key
    self.split_bbox = split_bbox

    # Load the detections
    self.labels_df = self.load_detection_files(
        detection_file_or_folder=detection_file_or_folder,
        detection_file_extension=detection_file_extension,
        image_folder=image_folder,
        use_absolute_filepaths=use_absolute_filepaths,
        strip_image_extension=strip_image_extension,
        image_path_key=image_path_key,
    )

    # Group the predictions
    self.grouped_labels_df = self.labels_df.groupby(by=self.image_path_key)

    # List the images
    self.image_names = list(self.grouped_labels_df.groups.keys())
    # Record the class names and number of classes
    self.class_names = np.unique(self.labels_df[self.label_key]).tolist()
    self.num_classes = len(self.class_names)

get_all_detections()

Return the aggregated detections dataframe

Source code in geograypher/predictors/derived_segmentors.py

def get_all_detections(self) -> pd.DataFrame:
    """Return the aggregated detections dataframe"""
    return self.labels_df

get_detection_centers(filename)

summary

Parameters:

Name	Type	Description	Default
filename	_type_	description	required

Returns:

Name	Type	Description
_type_		(n,2) array for (i,j) centers for each detection

Source code in geograypher/predictors/derived_segmentors.py

def get_detection_centers(self, filename):
    """_summary_

    Args:
        filename (_type_): _description_

    Returns:
        _type_: (n,2) array for (i,j) centers for each detection
    """
    if filename not in self.image_names:
        # Empty array of detection centers
        return np.zeros((0, 2))

    # Extract the corresponding dataframe
    df = self.grouped_labels_df.get_group(filename)

    all_corners = []
    for _, row in df.iterrows():
        corners = self.get_corners(row, as_int=False)
        all_corners.append(corners)

    all_corners = zip(*all_corners)
    all_corners = [np.array(x) for x in all_corners]

    imin, jmin, imax, jmax = all_corners

    # Average the left-right, top-bottom pairs
    centers = np.vstack([(imin + imax) / 2, (jmin + jmax) / 2]).T
    return centers

save_detection_data(output_csv_file)

Save the aggregated detections to a file

Parameters:

Name	Type	Description	Default
output_csv_file	PATH_TYPE	A path to a CSV file to save the detections to. The containing folder will be created if needed.	required

Source code in geograypher/predictors/derived_segmentors.py

def save_detection_data(self, output_csv_file: PATH_TYPE):
    """Save the aggregated detections to a file

    Args:
        output_csv_file (PATH_TYPE):
            A path to a CSV file to save the detections to. The containing folder will be
            created if needed.
    """
    ensure_containing_folder(output_csv_file)
    self.labels_df.to_csv(output_csv_file)