Function convert_verts_to_cam_coord

detrsmpl/utils/demo_utils.py:265–313

Convert vertices from the world coordinate to camera coordinate. Args: verts ([np.ndarray]): The vertices in the world coordinate. The shape is (frame,num_person,6890,3), (frame,6890,3), or (6890,3). pred_cams ([np.ndarray]): Camera parameters estimated b

(verts,
                               pred_cams,
                               bboxes_xy,
                               focal_length=5000.,
                               bbox_scale_factor=1.25,
                               bbox_format='xyxy')

def convert_verts_to_cam_coord(verts,
                               pred_cams,
                               bboxes_xy,
                               focal_length=5000.,
                               bbox_scale_factor=1.25,
                               bbox_format='xyxy'):
    """Convert vertices from the world coordinate to camera coordinate.

    Args:
        verts ([np.ndarray]): The vertices in the world coordinate.
            The shape is (frame,num_person,6890,3), (frame,6890,3),
            or (6890,3).
        pred_cams ([np.ndarray]): Camera parameters estimated by HMR or SPIN.
            The shape is (frame,num_person,3), (frame,3), or (3,).
        bboxes_xy ([np.ndarray]): (frame, num_person, 4|5), (frame, 4|5),
            or (4|5,)
        focal_length ([float],optional): Defined same as your training.
        bbox_scale_factor (float): scale factor for expanding the bbox.
        bbox_format (Literal['xyxy', 'xywh'] ): 'xyxy' means the left-up point
            and right-bottomn point of the bbox.
            'xywh' means the left-up point and the width and height of the
            bbox.
    Returns:
        np.ndarray: The vertices in the camera coordinate.
            The shape is (frame,num_person,6890,3) or (frame,6890,3).
        np.ndarray: The intrinsic parameters of the pred_cam.
            The shape is (num_frame, 3, 3).
    """
    K0 = get_default_hmr_intrinsic(focal_length=focal_length,
                                   det_height=224,
                                   det_width=224)
    K1 = convert_bbox_to_intrinsic(bboxes_xy,
                                   bbox_scale_factor=bbox_scale_factor,
                                   bbox_format=bbox_format)
    # K1K0(RX+T)-> K0(K0_inv K1K0)
    Ks = np.linalg.inv(K0) @ K1 @ K0
    # convert vertices from world to camera
    cam_trans = np.concatenate([
        pred_cams[..., [1]], pred_cams[..., [2]], 2 * focal_length /
        (224 * pred_cams[..., [0]] + 1e-9)
    ], -1)
    verts = verts + cam_trans[..., None, :]
    if verts.ndim == 4:
        verts = np.einsum('fnij,fnkj->fnki', Ks, verts)
    elif verts.ndim == 3:
        verts = np.einsum('fij,fkj->fki', Ks, verts)
    elif verts.ndim == 2:
        verts = np.einsum('fij,fkj->fki', Ks, verts[None])
    return verts, K0


def smooth_process(x,