Method convert_to_equirectangular

python/examples/panorama_sfm.py:307–427

Convert a reconstruction built from the rig of perspective virtual cameras back to one equirectangular camera/image per input panorama. The output reconstruction references the original panorama images with the native EQUIRECTANGULAR camera model. Frame poses, 3D points, and

(
        self, reconstruction: pycolmap.Reconstruction
    )

        raise ValueError(f"Unknown virtual camera for image {image_name!r}.")

    def convert_to_equirectangular(
        self, reconstruction: pycolmap.Reconstruction
    ) -> pycolmap.Reconstruction:
        """Convert a reconstruction built from the rig of perspective virtual
        cameras back to one equirectangular camera/image per input panorama.

        The output reconstruction references the original panorama images with
        the native EQUIRECTANGULAR camera model. Frame poses, 3D points, and
        all keypoints (including those without a 3D observation) are carried
        over by re-projecting the perspective keypoints onto the panorama
        through the same spherical mapping used for rendering, so the result is
        a valid, bundle-adjustable reconstruction.
        """
        if self._camera is None or self._pano_size is None:
            raise RuntimeError("No panorama was rendered yet.")
        pano_width, pano_height = self._pano_size

        equirect = pycolmap.Reconstruction()
        equirect_camera = pycolmap.Camera.create_from_model_id(
            camera_id=1,
            model=pycolmap.CameraModelId.EQUIRECTANGULAR,
            focal_length=0.0,
            width=pano_width,
            height=pano_height,
        )
        equirect.add_camera_with_trivial_rig(equirect_camera)

        # The rig reference sensor is virtual camera 0 (see
        # create_pano_rig_config), so rig_from_world == cam0_from_world and
        # pano_from_world = pano_from_cam0 @ cam0_from_world. The virtual
        # cameras share the panorama center, hence the zero translation.
        pano_from_ref = pycolmap.Rigid3d(
            pycolmap.Rotation3d(
                cast(NDArray3x3, self.cams_from_pano_rotation[0])
            ),
            cast(NDArray3x1, np.zeros((3, 1), dtype=np.float64)),
        ).inverse()

        # Group the registered virtual cameras by frame. All virtual cameras of
        # a frame observe the same panorama and share its pose, so we can
        # accumulate their keypoints into a single equirectangular image.
        images_by_frame: dict[int, list[pycolmap.Image]] = (
            collections.defaultdict(list)
        )
        for image in reconstruction.images.values():
            if image.has_pose:
                images_by_frame[image.frame_id].append(image)

        # Maps an image_id of a virtual camera to a dict from its old point2D
        # index to the (new point2D index, pano_name) in the equirectangular
        # image, so we can later rebuild the 3D point tracks.
        old_to_new_point2D: dict[int, dict[int, tuple[int, str]]] = {}
        pano_to_image_id: dict[str, int] = {}

        frame_images = sorted(
            images_by_frame.values(),
            key=lambda images: self.split_image_name(images[0].name)[1],
        )