opensfm/synthetic_data/synthetic_metrics.py

from typing import Tuple, List, Dict

import cv2
import numpy as np
import opensfm.transformations as tf
from opensfm import align, types, pymap, multiview


def points_errors(
    reference: types.Reconstruction, candidate: types.Reconstruction
) -> np.ndarray:
    common_points = set(reference.points.keys()).intersection(
        set(candidate.points.keys())
    )

    return np.array(
        [
            reference.points[p].coordinates - candidate.points[p].coordinates
            for p in common_points
        ]
    )


def completeness_errors(
    reference: types.Reconstruction, candidate: types.Reconstruction
) -> Tuple[float, float]:
    return (
        float(len(candidate.shots)) / float(len(reference.shots)),
        float(len(candidate.points)) / float(len(reference.points)),
    )


def gps_errors(candidate: types.Reconstruction) -> np.ndarray:
    errors = []
    for shot in candidate.shots.values():
        bias = candidate.biases[shot.camera.id]
        pose1 = bias.transform(shot.metadata.gps_position.value)
        pose2 = shot.pose.get_origin()
        errors.append(pose1 - pose2)
    return np.array(errors)


def gcp_errors(
    candidate: types.Reconstruction, gcps: Dict[str, pymap.GroundControlPoint]
) -> np.ndarray:
    errors = []
    for gcp in gcps.values():
        if not gcp.lla:
            continue

        triangulated = multiview.triangulate_gcp(gcp, candidate.shots, 1.0, 0.1)
        if triangulated is None:
            continue

        gcp_enu = candidate.reference.to_topocentric(*gcp.lla_vec)
        errors.append(triangulated - gcp_enu)
    return np.array(errors)


def position_errors(
    reference: types.Reconstruction, candidate: types.Reconstruction
) -> np.ndarray:
    common_shots = set(reference.shots.keys()).intersection(set(candidate.shots.keys()))
    errors = []
    for s in common_shots:
        pose1 = reference.shots[s].pose.get_origin()
        pose2 = candidate.shots[s].pose.get_origin()
        errors.append(pose1 - pose2)
    return np.array(errors)


def rotation_errors(
    reference: types.Reconstruction, candidate: types.Reconstruction
) -> np.ndarray:
    common_shots = set(reference.shots.keys()).intersection(set(candidate.shots.keys()))
    errors = []
    for s in common_shots:
        pose1 = reference.shots[s].pose.get_rotation_matrix()
        pose2 = candidate.shots[s].pose.get_rotation_matrix()
        difference = np.transpose(pose1).dot(pose2)
        rodrigues = cv2.Rodrigues(difference)[0].ravel()
        angle = np.linalg.norm(rodrigues)
        errors.append(angle)
    return np.array(errors)


def find_alignment(
    points0: List[np.ndarray], points1: List[np.ndarray]
) -> Tuple[float, np.ndarray, np.ndarray]:
    """Compute similarity transform between point sets.

    Returns (s, A, b) such that ``points1 = s * A * points0 + b``
    """
    v0, v1 = [], []
    for p0, p1 in zip(points0, points1):
        if p0 is not None and p1 is not None:
            v0.append(p0)
            v1.append(p1)
    v0 = np.array(v0).T
    v1 = np.array(v1).T
    M = tf.affine_matrix_from_points(v0, v1, shear=False)
    s = np.linalg.det(M[:3, :3]) ** (1.0 / 3.0)
    A = M[:3, :3] / s
    b = M[:3, 3]
    return s, A, b


def aligned_to_reference(
    reference: types.Reconstruction, reconstruction: types.Reconstruction
) -> types.Reconstruction:
    """Align a reconstruction to a reference."""
    coords1, coords2 = [], []
    for point1 in reconstruction.points.values():
        point2 = reference.points.get(point1.id)
        if point2 is not None:
            coords1.append(point1.coordinates)
            coords2.append(point2.coordinates)

    if len(coords1) == 0 or len(coords2) == 0:
        for shot1 in reconstruction.shots.values():
            shot2 = reference.shots.get(shot1.id)
            if shot2 is not None:
                coords1.append(shot1.pose.get_origin())
                coords2.append(shot2.pose.get_origin())

    s, A, b = find_alignment(coords1, coords2)
    aligned = _copy_reconstruction(reconstruction)
    align.apply_similarity(aligned, s, A, b)
    return aligned


def _copy_reconstruction(reconstruction: types.Reconstruction) -> types.Reconstruction:
    copy = types.Reconstruction()
    for camera in reconstruction.cameras.values():
        copy.add_camera(camera)
    for shot in reconstruction.shots.values():
        copy.add_shot(shot)
    for point in reconstruction.points.values():
        copy.add_point(point)
    return copy


def rmse(errors: np.ndarray) -> float:
    return np.sqrt(np.mean(errors ** 2))


def mad(errors: np.ndarray) -> float:
    return np.median(np.absolute(errors - np.median(errors)))