mrzmini 0.1.0__tar.gz

mrzmini-0.1.0/.gitignore

@@ -0,0 +1,14 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+build/
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+
+# Sample images used only by the parity harness
+testdata/
+uv.lock

mrzmini-0.1.0/LICENSE

@@ -0,0 +1,30 @@
+MIT License
+
+Copyright (c) 2026 Andrianina Rabakoson
+
+Portions of this software (the MRZ text parser in ``mrzmini/text.py`` and the
+pipeline engine in ``mrzmini/pipeline.py``) are copied verbatim from PassportEye
+and are:
+
+Copyright (c) 2016 Konstantin Tretyakov
+
+PassportEye is distributed under the MIT License. Its original copyright notice
+is preserved above and in the docstrings of the affected modules.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mrzmini-0.1.0/PKG-INFO

@@ -0,0 +1,137 @@
+Metadata-Version: 2.4
+Name: mrzmini
+Version: 0.1.0
+Summary: Minimal-footprint reimplementation of PassportEye's MRZ detection/reading (numpy + pillow only)
+Project-URL: Homepage, https://github.com/rbaks/mrzmini
+Project-URL: Repository, https://github.com/rbaks/mrzmini
+Project-URL: Issues, https://github.com/rbaks/mrzmini/issues
+Author-email: Andrianina Rabakoson <andri.bakoson@gmail.com>
+License-Expression: MIT
+License-File: LICENSE
+Keywords: id-document,mrz,ocr,passport,passporteye,tesseract
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Image Recognition
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Requires-Python: >=3.12
+Requires-Dist: numpy>=1.24
+Requires-Dist: pillow>=9
+Description-Content-Type: text/markdown
+
+# mrzmini
+
+A minimal-footprint, drop-in reimplementation of [PassportEye](https://github.com/konstantint/PassportEye)'s
+MRZ (machine-readable zone) detection and reading pipeline.
+
+PassportEye does one small, precise job — find and read the MRZ on an ID document —
+but drags in a heavy dependency stack to do it: `numpy`, `scipy`, `scikit-image`,
+`scikit-learn`, `matplotlib`, `imageio`, `pdfminer`, `pytesseract`.
+
+`mrzmini` reproduces the **exact same algorithm and output** using only:
+
+- **numpy**
+- **Pillow**
+- the external **`tesseract`** binary (called via `subprocess`; required only for OCR)
+
+Everything `skimage` / `sklearn` / `scipy` / `imageio` / `pdfminer` / `pytesseract`
+did is reimplemented on numpy in [`mrzmini/imageproc.py`](mrzmini/imageproc.py),
+[`mrzmini/geometry.py`](mrzmini/geometry.py), [`mrzmini/ocr.py`](mrzmini/ocr.py)
+and [`mrzmini/pdf.py`](mrzmini/pdf.py). The MRZ text parser
+([`text.py`](mrzmini/text.py)) and the pipeline engine ([`pipeline.py`](mrzmini/pipeline.py))
+are pure-Python and copied verbatim from upstream (MIT).
+
+## Install
+
+```bash
+pip install mrzmini
+```
+
+`mrzmini` needs the external **`tesseract`** OCR binary on `PATH` for the OCR step
+(detection alone does not need it):
+
+```bash
+apt install tesseract-ocr        # Debian/Ubuntu
+brew install tesseract           # macOS
+```
+
+Override the binary location with the `TESSERACT_CMD` environment variable.
+
+## Usage
+
+```python
+from mrzmini import read_mrz
+
+mrz = read_mrz('passport.jpg')        # path, bytes, file-like, or a .pdf
+print(mrz)                            # None if nothing was found
+print(mrz.to_dict())
+```
+
+The public surface mirrors PassportEye: `read_mrz(file, save_roi=False,
+extra_cmdline_params='')` returns an `MRZ` object with the same fields
+(`mrz_type`, `valid`, `valid_score`, `number`, `names`, `surname`,
+`date_of_birth`, `expiration_date`, `nationality`, `sex`, check digits, …).
+
+Command-line demo:
+
+```bash
+mrzmini passport.jpg                  # installed console script
+python -m mrzmini passport.jpg        # or via the module
+```
+
+## Requirements
+
+- Python ≥ 3.12, `numpy`, `pillow` (installed via `uv sync`).
+- The `tesseract` OCR binary on `PATH` (e.g. `apt install tesseract-ocr`).
+  Override its location with the `TESSERACT_CMD` environment variable.
+
+## How it works
+
+The pipeline is identical to PassportEye's (a lazy DAG of components):
+
+| Step | What it does | Upstream dependency replaced |
+|------|--------------|------------------------------|
+| `Loader` | read image → grayscale (color → float64 luma, gray → uint8) | `skimage.io` / `imageio`; `pdfminer` for PDFs |
+| `Scaler` | downscale so width ≤ 250 (anti-aliased) | `skimage.transform.rescale` (→ scipy) |
+| `BooneTransform` | `threshold_otsu(closing(\|sobel_v(black_tophat(img))\|))` → binary | `skimage.morphology` + `skimage.filters` |
+| `MRZBoxLocator` | marching-squares contours → `RotatedBox` (PCA) → merge parallel boxes | `skimage.measure` + `sklearn.PCA` |
+| `extract_from_image` | un-rotate + crop the ROI | `skimage.transform.rotate` |
+| `ocr` | run Tesseract on the ROI | `pytesseract` |
+| `MRZ.from_ocr` | clean up + parse + checksum-validate | (pure Python) |
+
+## Parity with PassportEye
+
+[`parity_check.py`](parity_check.py) compares `mrzmini` against the real
+PassportEye stage by stage. Across the **entire PassportEye test corpus
+(36 images: TD1/TD2/TD3/MRVA/MRVB, scores 0–100, JPG/PNG/PDF)**:
+
+- `img_binary`: **0 pixel disagreements** on every image
+- detected MRZ **boxes: identical** on every image
+- `read_mrz(...).to_dict()`: **identical** on every image
+
+```bash
+uv run --group parity python parity_check.py            # scans testdata/
+uv run --group parity python parity_check.py some.jpg   # specific files
+```
+
+Internally every reimplemented primitive matches its scikit-image / scipy /
+scikit-learn counterpart to floating-point precision (bilinear resize and
+marching-squares contours are bit-exact; the bicubic OCR-retry resize matches
+scipy to ~1e-14, and the PCA box geometry to ~1e-12).
+
+## Notes / limitations
+
+- PDF support is best-effort: like upstream it extracts the first embedded JFIF
+  JPEG (`\xff\xd8\xff\xe0`); other embedded image encodings are not handled.
+- Tesseract is an external binary, not a Python library, so it remains a
+  requirement for the OCR step (detection alone does not need it).
+
+## License
+
+The reused PassportEye algorithm and copied modules are MIT
+(© Konstantin Tretyakov). Reimplemented primitives follow scikit-image /
+scipy behavior (BSD).

mrzmini-0.1.0/README.md

@@ -0,0 +1,112 @@
+# mrzmini
+
+A minimal-footprint, drop-in reimplementation of [PassportEye](https://github.com/konstantint/PassportEye)'s
+MRZ (machine-readable zone) detection and reading pipeline.
+
+PassportEye does one small, precise job — find and read the MRZ on an ID document —
+but drags in a heavy dependency stack to do it: `numpy`, `scipy`, `scikit-image`,
+`scikit-learn`, `matplotlib`, `imageio`, `pdfminer`, `pytesseract`.
+
+`mrzmini` reproduces the **exact same algorithm and output** using only:
+
+- **numpy**
+- **Pillow**
+- the external **`tesseract`** binary (called via `subprocess`; required only for OCR)
+
+Everything `skimage` / `sklearn` / `scipy` / `imageio` / `pdfminer` / `pytesseract`
+did is reimplemented on numpy in [`mrzmini/imageproc.py`](mrzmini/imageproc.py),
+[`mrzmini/geometry.py`](mrzmini/geometry.py), [`mrzmini/ocr.py`](mrzmini/ocr.py)
+and [`mrzmini/pdf.py`](mrzmini/pdf.py). The MRZ text parser
+([`text.py`](mrzmini/text.py)) and the pipeline engine ([`pipeline.py`](mrzmini/pipeline.py))
+are pure-Python and copied verbatim from upstream (MIT).
+
+## Install
+
+```bash
+pip install mrzmini
+```
+
+`mrzmini` needs the external **`tesseract`** OCR binary on `PATH` for the OCR step
+(detection alone does not need it):
+
+```bash
+apt install tesseract-ocr        # Debian/Ubuntu
+brew install tesseract           # macOS
+```
+
+Override the binary location with the `TESSERACT_CMD` environment variable.
+
+## Usage
+
+```python
+from mrzmini import read_mrz
+
+mrz = read_mrz('passport.jpg')        # path, bytes, file-like, or a .pdf
+print(mrz)                            # None if nothing was found
+print(mrz.to_dict())
+```
+
+The public surface mirrors PassportEye: `read_mrz(file, save_roi=False,
+extra_cmdline_params='')` returns an `MRZ` object with the same fields
+(`mrz_type`, `valid`, `valid_score`, `number`, `names`, `surname`,
+`date_of_birth`, `expiration_date`, `nationality`, `sex`, check digits, …).
+
+Command-line demo:
+
+```bash
+mrzmini passport.jpg                  # installed console script
+python -m mrzmini passport.jpg        # or via the module
+```
+
+## Requirements
+
+- Python ≥ 3.12, `numpy`, `pillow` (installed via `uv sync`).
+- The `tesseract` OCR binary on `PATH` (e.g. `apt install tesseract-ocr`).
+  Override its location with the `TESSERACT_CMD` environment variable.
+
+## How it works
+
+The pipeline is identical to PassportEye's (a lazy DAG of components):
+
+| Step | What it does | Upstream dependency replaced |
+|------|--------------|------------------------------|
+| `Loader` | read image → grayscale (color → float64 luma, gray → uint8) | `skimage.io` / `imageio`; `pdfminer` for PDFs |
+| `Scaler` | downscale so width ≤ 250 (anti-aliased) | `skimage.transform.rescale` (→ scipy) |
+| `BooneTransform` | `threshold_otsu(closing(\|sobel_v(black_tophat(img))\|))` → binary | `skimage.morphology` + `skimage.filters` |
+| `MRZBoxLocator` | marching-squares contours → `RotatedBox` (PCA) → merge parallel boxes | `skimage.measure` + `sklearn.PCA` |
+| `extract_from_image` | un-rotate + crop the ROI | `skimage.transform.rotate` |
+| `ocr` | run Tesseract on the ROI | `pytesseract` |
+| `MRZ.from_ocr` | clean up + parse + checksum-validate | (pure Python) |
+
+## Parity with PassportEye
+
+[`parity_check.py`](parity_check.py) compares `mrzmini` against the real
+PassportEye stage by stage. Across the **entire PassportEye test corpus
+(36 images: TD1/TD2/TD3/MRVA/MRVB, scores 0–100, JPG/PNG/PDF)**:
+
+- `img_binary`: **0 pixel disagreements** on every image
+- detected MRZ **boxes: identical** on every image
+- `read_mrz(...).to_dict()`: **identical** on every image
+
+```bash
+uv run --group parity python parity_check.py            # scans testdata/
+uv run --group parity python parity_check.py some.jpg   # specific files
+```
+
+Internally every reimplemented primitive matches its scikit-image / scipy /
+scikit-learn counterpart to floating-point precision (bilinear resize and
+marching-squares contours are bit-exact; the bicubic OCR-retry resize matches
+scipy to ~1e-14, and the PCA box geometry to ~1e-12).
+
+## Notes / limitations
+
+- PDF support is best-effort: like upstream it extracts the first embedded JFIF
+  JPEG (`\xff\xd8\xff\xe0`); other embedded image encodings are not handled.
+- Tesseract is an external binary, not a Python library, so it remains a
+  requirement for the OCR step (detection alone does not need it).
+
+## License
+
+The reused PassportEye algorithm and copied modules are MIT
+(© Konstantin Tretyakov). Reimplemented primitives follow scikit-image /
+scipy behavior (BSD).

mrzmini-0.1.0/mrzmini/__init__.py

@@ -0,0 +1,17 @@
+'''
+mrzmini: a minimal-footprint reimplementation of PassportEye's MRZ
+detection and reading pipeline.
+
+Only depends on numpy + Pillow (plus the external ``tesseract`` binary for OCR).
+The public interface mirrors PassportEye:
+
+    from mrzmini import read_mrz
+    mrz = read_mrz('passport.jpg')
+    print(mrz.to_dict())
+
+Original algorithm and code: Konstantin Tretyakov, PassportEye (MIT License).
+'''
+from .text import MRZ
+from .image import read_mrz, MRZPipeline
+
+__all__ = ['read_mrz', 'MRZ', 'MRZPipeline']

mrzmini-0.1.0/mrzmini/__main__.py

@@ -0,0 +1,28 @@
+"""Command-line entry point for mrzmini.
+
+Usage:
+    mrzmini <image-or-pdf> [...]
+    python -m mrzmini <image-or-pdf> [...]
+"""
+import sys
+import json
+
+from mrzmini import read_mrz
+
+
+def main():
+    if len(sys.argv) < 2:
+        print(__doc__)
+        return
+    for path in sys.argv[1:]:
+        print("=" * 60)
+        print(path)
+        mrz = read_mrz(path)
+        if mrz is None:
+            print("  No MRZ found.")
+            continue
+        print(json.dumps(mrz.to_dict(), indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()

mrzmini-0.1.0/mrzmini/geometry.py

@@ -0,0 +1,241 @@
+'''
+mrzmini.geometry: RotatedBox geometry, adapted from PassportEye's
+``passporteye/util/geometry.py``.
+
+Changes vs upstream (to drop heavy dependencies), behaviour preserved:
+  * ``sklearn.decomposition.PCA`` -> a small numpy SVD (see :func:`_pca`).
+    The resulting box (center/width/height/angle) is invariant to the sign of
+    the principal components, so sklearn's ``svd_flip`` convention need not be
+    reproduced.
+  * ``skimage.transform`` (``rotate`` / ``SimilarityTransform``) ->
+    :mod:`mrzmini.imageproc`.
+  * The matplotlib ``plot`` method is removed.
+
+Author of original: Konstantin Tretyakov. License: MIT.
+'''
+import numpy as np
+
+from .imageproc import rotate as _rotate, SimilarityTransform
+
+
+def _pca(points):
+    """Return ``(mean, components)`` matching ``sklearn.decomposition.PCA(2).fit``.
+
+    ``components`` rows are the principal axes ordered by descending variance
+    (row 0 = longest axis). ``transform(x) = (x - mean) @ components.T``.
+    """
+    mean = points.mean(axis=0)
+    centered = points - mean
+    # full_matrices=False -> Vt has shape (min(n,2), 2); singular values descending.
+    _u, _s, vt = np.linalg.svd(centered, full_matrices=False)
+    components = vt
+    if components.shape[0] < 2:
+        # Degenerate (all points collinear with <2 dims of variance): complete
+        # the basis with an orthogonal vector so downstream maths is well-defined.
+        v0 = components[0]
+        components = np.vstack([v0, [-v0[1], v0[0]]])
+    return mean, components
+
+
+class RotatedBox(object):
+    """
+    RotatedBox represents a rectangular box centered at (cx,cy) with dimensions width x height,
+    rotated by angle radians counterclockwise.
+
+    >>> RotatedBox.from_points([[0,0], [2,1], [0,1], [2,0]])
+    RotatedBox(cx=1.0, cy=0.5, width=2.0, height=1.0, angle=0.0)
+    """
+
+    def __init__(self, center, width, height, angle, points=None):
+        """Creates a new RotatedBox.
+
+        :param points: This parameter may be used to indicate the set of points used to create the box.
+        """
+        self.center = np.asarray(center, dtype=np.float64)
+        self.width = width
+        self.height = height
+        self.angle = angle
+        self.points = points
+
+    def __repr__(self):
+        return "RotatedBox(cx={0}, cy={1}, width={2}, height={3}, angle={4})".format(self.cx, self.cy, self.width, self.height, self.angle)
+
+    @property
+    def cx(self):
+        return self.center[0]
+
+    @property
+    def cy(self):
+        return self.center[1]
+
+    @property
+    def area(self):
+        return self.width * self.height
+
+    def approx_equal(self, center, width, height, angle, tol=1e-6):
+        "Method mainly useful for testing"
+        return abs(self.cx - center[0]) < tol and abs(self.cy - center[1]) < tol and abs(self.width - width) < tol and \
+               abs(self.height - height) < tol and abs(self.angle - angle) < tol
+
+    def rotated(self, rotation_center, angle):
+        """Returns a RotatedBox that is obtained by rotating this box around a given center by a given angle.
+
+        >>> assert RotatedBox([2, 2], 2, 1, 0.1).rotated([1, 1], np.pi/2).approx_equal([0, 2], 2, 1, np.pi/2+0.1)
+        """
+        rot = np.array([[np.cos(angle), np.sin(angle)], [-np.sin(angle), np.cos(angle)]])
+        t = np.asarray(rotation_center, dtype=np.float64)
+        new_c = np.dot(rot.T, (self.center - t)) + t
+        return RotatedBox(new_c, self.width, self.height, (self.angle+angle) % (np.pi*2))
+
+    def as_poly(self, margin_width=0, margin_height=0):
+        """Converts this box to a polygon, i.e. 4x2 array, representing the four corners starting from lower left to upper left counterclockwise.
+
+        :param margin_width: The additional "margin" that will be added to the box along its width dimension (from both sides) before conversion.
+        :param margin_height: The additional "margin" that will be added to the box along its height dimension (from both sides) before conversion.
+
+        >>> RotatedBox([0, 0], 4, 2, 0).as_poly()
+        array([[-2., -1.],
+               [ 2., -1.],
+               [ 2.,  1.],
+               [-2.,  1.]])
+        >>> RotatedBox([0, 0], 4, 2, np.pi/2).as_poly()
+        array([[ 1., -2.],
+               [ 1.,  2.],
+               [-1.,  2.],
+               [-1., -2.]])
+        >>> RotatedBox([0, 0], 0, 0, np.pi/2).as_poly(2, 1)
+        array([[ 1., -2.],
+               [ 1.,  2.],
+               [-1.,  2.],
+               [-1., -2.]])
+        """
+        v_hor = (self.width/2 + margin_width)*np.array([np.cos(self.angle), np.sin(self.angle)])
+        v_vert = (self.height/2 + margin_height)*np.array([-np.sin(self.angle), np.cos(self.angle)])
+        c = np.array([self.cx, self.cy])
+        return np.vstack([c - v_hor - v_vert, c + v_hor - v_vert, c + v_hor + v_vert, c - v_hor + v_vert])
+
+    def extract_from_image(self, img, scale=1.0, margin_width=5, margin_height=5):
+        """Extracts the contents of this box from a given image.
+        For that the image is "unrotated" by the appropriate angle, and the corresponding part is extracted from it.
+
+        Returns an image with dimensions height*scale x width*scale.
+        Note that the box coordinates are interpreted as "image coordinates" (i.e. x is row and y is column),
+        and box angle is considered to be relative to the vertical (i.e. np.pi/2 is "normal orientation")
+
+        :param img: a numpy ndarray suitable for image processing.
+        :param scale: the RotatedBox is scaled by this value before performing the extraction.
+            This is necessary when, for example, the location of a particular feature is determined using a smaller image,
+            yet then the corresponding area needs to be extracted from the original, larger image.
+            The scale parameter in this case should be width_of_larger_image/width_of_smaller_image.
+        :param margin_width: The margin that should be added to the width dimension of the box from each size.
+            This value is given wrt actual box dimensions (i.e. not scaled).
+        :param margin_height: The margin that should be added to the height dimension of the box from each side.
+        :return: a numpy ndarray, corresponding to the extracted region (aligned straight).
+        """
+        rotate_by = (np.pi/2 - self.angle)*180/np.pi
+        img_rotated = _rotate(img, angle=rotate_by, center=[self.center[1]*scale, self.center[0]*scale], resize=True)
+        # The resizeable transform will shift the resulting image somewhat wrt original coordinates.
+        # When we cut out the box we will compensate for this shift.
+        shift_c, shift_r = self._compensate_rotation_shift(img, scale)
+
+        r1 = max(int((self.center[0] - self.height/2 - margin_height)*scale - shift_r), 0)
+        r2 = int((self.center[0] + self.height/2 + margin_height)*scale - shift_r)
+        c1 = max(int((self.center[1] - self.width/2 - margin_width)*scale - shift_c), 0)
+        c2 = int((self.center[1] + self.width/2 + margin_width)*scale - shift_c)
+        return img_rotated[r1:r2, c1:c2]
+
+    def _compensate_rotation_shift(self, img, scale):
+        """This is an auxiliary method used by extract_from_image.
+        It repeats the shift computation performed by skimage's resize-enabled rotate,
+        so that the box can be cut out at the right offset.
+        """
+        ctr = np.asarray([self.center[1]*scale, self.center[0]*scale])
+        tform1 = SimilarityTransform(translation=ctr)
+        tform2 = SimilarityTransform(rotation=np.pi/2 - self.angle)
+        tform3 = SimilarityTransform(translation=-ctr)
+        tform = tform3 + tform2 + tform1
+
+        rows, cols = img.shape[0], img.shape[1]
+        corners = np.array([
+            [0, 0],
+            [0, rows - 1],
+            [cols - 1, rows - 1],
+            [cols - 1, 0]
+        ])
+        corners = tform.inverse(corners)
+        minc = corners[:, 0].min()
+        minr = corners[:, 1].min()
+        maxc = corners[:, 0].max()
+        maxr = corners[:, 1].max()
+
+        # fit output image in new shape
+        translation = (minc, minr)
+        tform4 = SimilarityTransform(translation=translation)
+        tform = tform4 + tform
+        tform.params[2] = (0, 0, 1)
+
+        # Compute the shift of the transformed center wrt original
+        return (ctr - tform.inverse(ctr)).ravel().tolist()
+
+    @staticmethod
+    def from_points(points, box_type='bb'):
+        """
+        Interpret a given point cloud as a RotatedBox, using PCA to determine the potential orientation (the longest component becomes width)
+        This is basically an approximate version of a min-area-rectangle algorithm.
+
+        :param points: An n x 2 numpy array of coordinates.
+        :param box_type: The kind of method used to estimate the "box".
+            Possible values:
+                - `'bb'`, denoting the "bounding box" approach (min/max coordinates of the points correspond to box limits)
+                - `'mrz`, denoting a slightly modified technique, suited for MRZ zone detection from contour images.
+                          Here the assumption is that the upper and lower bounds of the box are better estimated as the
+                          10% and 90% quantile of the corresponding coordinates (rather than 0% and 100%, i.e. min and max).
+                          This helps against accidental noise in the contour.
+                          The `'mrz'` correction is only applied when there are at least 10 points in the set.
+        :returns: a RotatedBox, bounding the given set of points, oriented according to the principal components.
+
+        >>> RotatedBox.from_points([[0,0]])
+        RotatedBox(cx=0.0, cy=0.0, width=0.0, height=0.0, angle=0.0)
+        >>> assert RotatedBox.from_points([[0,0], [1,1], [2,2]]).approx_equal([1, 1], np.sqrt(8), 0, np.pi/4)
+        >>> assert RotatedBox.from_points([[0,0], [1,1], [0,1], [1,0]]).approx_equal([0.5, 0.5], 1, 1, 0.0) # The angle is rather arbitrary here
+        >>> assert RotatedBox.from_points([[0,0], [2,1], [0,1], [2,0]]).approx_equal([1, 0.5], 2, 1, 0)
+        >>> assert RotatedBox.from_points([[0,0], [2,4], [0,4], [2,0]]).approx_equal([1, 2], 4, 2, np.pi/2)
+        >>> assert RotatedBox.from_points([[0,0], [1,1.5], [2,0]]).approx_equal([1, 0.75], 2, 1.5, 0)
+        >>> assert RotatedBox.from_points([[0,0], [0,1], [1,1]]).approx_equal([0.25, 0.75], np.sqrt(2), np.sqrt(2)/2, np.pi/4)
+        """
+        points = np.asarray(points, dtype=np.float64)
+        if points.shape[0] == 1:
+            return RotatedBox(points[0], width=0.0, height=0.0, angle=0.0, points=points)
+
+        mean, components = _pca(points)
+        # Find the angle
+        angle = (np.arctan2(components[0, 1], components[0, 0]) % np.pi)
+        if abs(angle - np.pi) < angle:
+            # Here the angle is always between -pi and pi
+            # If the principal component happened to be oriented so that the angle happens to be > pi/2 by absolute value,
+            # we flip the direction
+            angle = angle - np.pi if angle > 0 else angle + np.pi
+        points_transformed = (points - mean) @ components.T
+        ll = np.min(points_transformed, 0)
+        ur = np.max(points_transformed, 0)
+        wh = ur - ll
+
+        # Now compute and return the bounding box
+        if box_type == 'bb' or (box_type == 'mrz' and points.shape[0] < 10):
+            # We know that if we rotate the points around mean, we get a box with bounds ur and ll
+            # The center of this box is (ur+ll)/2 + mean, which is not the same as the mean,
+            # hence to get the center of the original box we need to "unrotate" this box back.
+            return RotatedBox(np.dot(components.T, (ll+ur)/2) + mean, width=wh[0], height=wh[1], angle=angle, points=points)
+        elif box_type == 'mrz':
+            # When working with MRZ detection from contours, we may have minor "bumps" in the contour,
+            # that should be ignored at least along the long ("horizontal") side.
+            # To do that, we will use 10% and 90% quantiles as the bounds of the box instead of the max and min.
+            h_coord = sorted(points_transformed[:, 1])
+            n = len(h_coord)
+            bottom, top = h_coord[n // 10], h_coord[n * 9 // 10]
+            valid_points = np.logical_and(points_transformed[:, 1] >= bottom, points_transformed[:, 1] <= top)
+            rb = RotatedBox.from_points(points[valid_points, :], 'bb')
+            rb.points = points
+            return rb
+        else:
+            raise ValueError("Unknown parameter value: box_type=%s" % box_type)