Greedy IoU matching: highest IoU pairs are matched first.

Scale the bounding box by ``sx`` and ``sy``.

Character error rate between two strings via edit distance.

Loads a document from a given file path based on the file extension. This function determines the appropriate document type to load by e

Retrieve the value of an environment variable. Args: name (str): The name of the environment variable to retrieve. default (

Cluster aabbs with DBSCAN on the Jaccard distance between boxes.

Return the area of the bounding box.

Intersection over union of two axis-aligned rectangles.

Normalise a grayscale image to roughly ``[-0.5, 0.5]``.

Save document page as image. #TODO: I am using `matplotlib` here. Alternatively, OpenCV could do the trick as well. :param

Translate the bounding box by ``(dx, dy)``.

Run HTR on a document and return word-level predictions. Bounding box coordinates are always in document units (72 DPI), regardless of the in

Export ``checkpoint`` to ``output_dir`` as ``model.onnx`` + ``config.json``.

Main function that performs HTR. This function exports an Xournal(++) file to a PDF file, performs Handwritten Text Recognition (HTR) on the file

Take a max number of pixels per connected component (>=3 for DBSCAN).

Resolve a device string, with auto-detection for ``"auto"``.

Parse arguments from command line.

TODO. These are the steps performed: 1. Loop over `self.positions` to obtain index and location. 2. Get sample from dataset.

Detect word bounding boxes in a grayscale image. The returned boxes are in the pixel coordinate system of the *passed* image (the int

Draw bounding boxes on a (possibly normalised, grayscale) image.

Export a Xournal(++) file to PDF using Xournal++. This function uses the `xournalpp` command-line tool to convert a Xournal(++) file specifie

Resolve a device string, with auto-detection for ``"auto"``.

Generate a file name based on an index and a file type. This method returns a string in the format: `"{file_type}_{idx:06}.png"`,

Best-path CTC decoding. Args: log_probs: (seq_len, batch, num_classes) from ``forward()``. charset: list of characters (blank ind

Return path to the IAM-DB dataset, downloading from HuggingFace Hub if needed. The IAM Handwriting Database (offline forms) is stored as raw file

Return benchmark samples from the xournalpp_htr_benchmark HuggingFace dataset.

Load ``ModelConfig`` from a training run's ``config.yaml``.

Recognise text in a grayscale word image. The image is resized to the network's expected input dimensions (uniform scale, centered on

TODO. Steps that are performed: TODO. TODO: Determine page sizes etc & adjust rendering

Restrict foreground indices to a maximum number.

Writes handwritten text predictions to a PDF file. This function reads an input PDF file using PyMuPDF, extracts each page from the PDF, and

Visualise the CTC output matrix, showing only the top-k characters.