Method __init__

detrsmpl/core/cameras/cameras.py:22–129

Initialize your cameras with `build_cameras` following: 1): provide `K`, `R`, `T`, `resolution`/`image_size`, `in_ndc` directly. `K` should be shape of (N, 3, 3) or (N, 4, 4). `R` should be shape of (N, 3, 3). `T` should be shape of (N, 3).

(self, **kwargs)

class MMCamerasBase(cameras.CamerasBase):
    """Inherited from Pytorch3D CamerasBase and provide some new functions."""
    def __init__(self, **kwargs) -> None:
        """Initialize your cameras with `build_cameras` following:

        1): provide `K`, `R`, `T`, `resolution`/`image_size`, `in_ndc`
            directly.
            `K` should be shape of (N, 3, 3) or (N, 4, 4).
            `R` should be shape of (N, 3, 3).
            `T` should be shape of (N, 3).
        2): if `K` is not provided, will use `get_default_projection_matrix`
            to generate K from camera intrinsic parameters.
            E.g., you can pass `focal_length`, `principal_point` for
            perspective camers.
            If these args are not provided, will use default values.
        3): if `R` is not provided, will use Identity matrix as default.
        4): if `T` is not provided, will use zeros matrix as default.
        5): `convention` means your source parameter camera convention.
            This mainly depends on how you get the matrixs. E.g., you get the
            `K` `R`, `T` by calibration with opencv, you should set
            `convention = opencv`. To figure out your camera convention,
            please see the definition of its extrinsic and intrinsic matrixs.
            For projection and rendering, the matrixs will be converted to
            `pytorch3d` finally since the `transforms3d` called in rendering
            and projection are defined as `pytorch3d` convention.
        6): `image_size` equals `resolution`.
        7): `in_ndc` could be set for 'PerspectiveCameras' and
            'OrthographicCameras', other cameras are fixed for this arg.
            `in_ndc = True` means your projection matrix is defined as `camera
            space to NDC space`. Under this cirecumstance you need to set
            `image_size` or `resolution` (they are equal) when you need to do
            `transform_points_screen`. You can also override resolution
            in `transform_points_screen` function.
            `in_ndc = False` means your projections matrix is defined as
            `cameras space to screen space`. Under this cirecumstance you do
            not need to set `image_size` or `resolution` (they are equal) when
            you need to do `transform_points_screen` since the projection
            matrix is defined as view space to screen space.
        """
        for k in kwargs:
            if isinstance(kwargs.get(k), np.ndarray):
                kwargs.update({k: torch.Tensor(kwargs[k])})
        convention = kwargs.pop('convention', 'pytorch3d').lower()
        in_ndc = kwargs.pop('in_ndc', kwargs.get('_in_ndc'))
        kwargs.update(_in_ndc=in_ndc)
        is_perspective = kwargs.get('_is_perspective')
        kwargs.pop('is_perspective', None)

        image_size = kwargs.get('image_size', kwargs.get('resolution', None))

        if image_size is not None:
            if isinstance(image_size, (int, float)):
                image_size = (image_size, image_size)
            if isinstance(image_size, (tuple, list)):
                image_size = torch.Tensor(image_size)
            if isinstance(image_size, torch.Tensor):
                if image_size.numel() == 1:
                    image_size = image_size.repeat(2)
                image_size = image_size.view(-1, 2)