miop.rectification

Classes

Rectification()

Performs stereo image rectification based on affine fundamental matrices.

class miop.rectification.Rectification

Bases: DAGNode

Performs stereo image rectification based on affine fundamental matrices.

This class computes rectifying transformations (similarity or rigid), applies them to stereo image pairs, and returns rectified image pairs along with transformation matrices.

rectification_transforms

List of transformation matrices for each image pair. Each dict contains ‘T1’, ‘T2’, and ‘Ht’.

Type:

list of dict

rectified_pairs

List of rectified image pairs, each as a 2-element NumPy array of images.

Type:

list of np.ndarray

eval(epipolar_sets, img_collection)

Applies rectifying transformations to each stereo image pair in a dataset.

Parameters:

• epipolar_sets (list of dict) – Each dict contains at least a fundamental matrix under key ‘F’.

• img_collection (object) – Object that holds image pairs in the pairs attribute. Each pair is a 2-tuple of image objects with .image fields.

Returns:

A tuple of:

• rectification_transformslist of dict

  Each dict contains matrices ‘T1’, ‘T2’, and ‘Ht’ (post-translation).

• rectified_pairslist of np.ndarray

  Each element is an array of two rectified images.

Return type:

tuple

homogenize(m)

Converts 2D point coordinates to homogeneous coordinates.

Parameters:

m (np.ndarray) – 2D coordinates, shape (n, 2) or (2, n)

Returns:

Homogeneous coordinates, shape (n, 3) or (3, n) to match input format.

Return type:

np.ndarray

rectifying_rigid_transformations(F, pts1, pts2)

Computes rigid rectifying transformations from an affine F and matched points.

Parameters:

• F (np.ndarray) – 3x3 affine fundamental matrix.

• pts1 (np.ndarray) – 3xn homogeneous point array from image 1.

• pts2 (np.ndarray) – 3xn homogeneous point array from image 2.

Returns:

S1, S2np.ndarray

Rigid rectifying transformation matrices for the two images.

Return type:

tuple

Raises:

• AssertionError – If F is not affine or not rank 2.

• Exception – If rotation matrix signs cannot be resolved.

rectifying_similarities(F)

Computes rectifying similarity transforms from an affine fundamental matrix.

Based on Shimshoni’s geometric interpretation of weak-perspective motion.

Parameters:

F (np.ndarray) – 3x3 affine fundamental matrix.

Returns:

S1, S2np.ndarray

Similarity transformations for image 1 and image 2.

Return type:

tuple

Raises:

AssertionError – If F is not affine or not rank 2 or not 3x3.

transform_2D(pts, H)

Applies a homography to 2D homogeneous points.

Parameters:

• pts (np.ndarray) – Homogeneous 2D points (3, n).

• H (np.ndarray) – Homography matrix (3, 3).

Returns:

Transformed 2D points (2, n).

Return type:

np.ndarray

transform_images(H1, H2, img1, img2)

Applies given homographies to images and aligns them using a shared bounding box.

Parameters:

• H1 (np.ndarray) – 3x3 homography for image 1.

• H2 (np.ndarray) – 3x3 homography for image 2.

• img1 (np.ndarray) – Original image 1.

• img2 (np.ndarray) – Original image 2.

Returns:

rectified_img1, rectified_img2np.ndarray

Warped images.

Htnp.ndarray

Translation matrix used to shift both images into shared frame.

Return type:

tuple

transform_matches(pts1, pts2, H1, H2)

Applies rectifying homographies to matched keypoints.

Parameters:

• pts1 (np.ndarray) – Matched keypoints from image 1, shape (n, 2).

• pts2 (np.ndarray) – Matched keypoints from image 2, shape (n, 2).

• H1 (np.ndarray) – Homography for image 1.

• H2 (np.ndarray) – Homography for image 2.

Returns:

pts1r, pts2rnp.ndarray

Rectified keypoints in both images, shape (n, 3).

Return type:

tuple