Method decode_single

monai/apps/detection/utils/box_coder.py:198–243

From a set of original boxes and encoded relative box offsets, Args: rel_codes: encoded boxes, Nx(4*num_box_reg) or Nx(6*num_box_reg) torch tensor. reference_boxes: reference boxes, Nx4 or Nx6 torch tensor. The box mode is assumed to be ``StandardMode``

(self, rel_codes: Tensor, reference_boxes: Tensor)

        return pred_boxes

    def decode_single(self, rel_codes: Tensor, reference_boxes: Tensor) -> Tensor:
        """
        From a set of original boxes and encoded relative box offsets,

        Args:
            rel_codes: encoded boxes, Nx(4*num_box_reg) or Nx(6*num_box_reg) torch tensor.
            reference_boxes: reference boxes, Nx4 or Nx6 torch tensor. The box mode is assumed to be ``StandardMode``

        Return:
            decoded boxes, Nx(4*num_box_reg) or Nx(6*num_box_reg) torch tensor. The box mode will to be ``StandardMode``
        """
        reference_boxes = reference_boxes.to(rel_codes.dtype)
        offset = reference_boxes.shape[-1]

        pred_boxes = []
        boxes_cccwhd: torch.Tensor = convert_box_mode(
            reference_boxes, src_mode=StandardMode, dst_mode=CenterSizeMode
        )  # type: ignore[assignment]

        for axis in range(self.spatial_dims):
            whd_axis = boxes_cccwhd[:, axis + self.spatial_dims]
            ctr_xyz_axis = boxes_cccwhd[:, axis]
            dxyz_axis = rel_codes[:, axis::offset] / self.weights[axis]
            dwhd_axis = rel_codes[:, self.spatial_dims + axis :: offset] / self.weights[axis + self.spatial_dims]
            # Prevent sending too large values into torch.exp()
            dwhd_axis = torch.clamp(dwhd_axis.to(COMPUTE_DTYPE), max=self.boxes_xform_clip)

            pred_ctr_xyx_axis = dxyz_axis * whd_axis[:, None] + ctr_xyz_axis[:, None]
            pred_whd_axis = torch.exp(dwhd_axis) * whd_axis[:, None]
            pred_whd_axis = pred_whd_axis.to(dxyz_axis.dtype)  # type: ignore[union-attr]

            # When convert float32 to float16, Inf or Nan may occur
            if torch.isnan(pred_whd_axis).any() or torch.isinf(pred_whd_axis).any():
                raise ValueError("pred_whd_axis is NaN or Inf.")

            # Distance from center to box's corner.
            c_to_c_whd_axis = (
                torch.tensor(0.5, dtype=pred_ctr_xyx_axis.dtype, device=pred_whd_axis.device) * pred_whd_axis  # type: ignore[arg-type]
            )

            pred_boxes.append(pred_ctr_xyx_axis - c_to_c_whd_axis)
            pred_boxes.append(pred_ctr_xyx_axis + c_to_c_whd_axis)

        pred_boxes = pred_boxes[::2] + pred_boxes[1::2]
        pred_boxes_final = torch.stack(pred_boxes, dim=2).flatten(1)
        return pred_boxes_final