def _get_best_yaw_obj_from_pos(self, obj_id, grid_pos, height=1.0, use_iou=True):
    obj = self.objects[obj_id]
    obj_fine_class = obj['fine_class']

    cx, cy = self.env.house.to_coor(grid_pos[0], grid_pos[1])
    self.env.cam.pos.x = cx
    self.env.cam.pos.y = height
    self.env.cam.pos.z = cy

    best_yaw, best_coverage, best_mask = None, 0, None
    for yaw in self.angles:
      self.env.cam.yaw = yaw
      self.env.cam.updateDirection()

      seg = self.env.render(mode='semantic')
      c = self.semantic_classes[obj_fine_class.lower()]
      mask = np.all(seg == c, axis=2)
      if use_iou:
        coverage = self._compute_iou(mask)
      else:
        coverage = np.sum(mask) / (seg.shape[0] * seg.shape[1])

      if best_yaw == None:
        best_yaw = yaw
        best_coverage = coverage
        best_mask = mask
      else:
        if coverage > best_coverage:
          best_yaw = yaw
          best_coverage = coverage
          best_mask = mask

    return best_yaw, best_coverage, best_mask

  def _compute_iou(self, cand_mask, ref_mask=None):
    """
    Given (h, w) cand_mask, we wanna our ref_mask to be in the center of image,
    with [0.25h:0.75h, 0.25w:0.75w] occupied.
    """
    if ref_mask is None:
        h, w = cand_mask.shape[0], cand_mask.shape[1]
        ref_mask = np.zeros((h,w), np.int8)
        ref_mask[int(0.25*h):int(0.85*h), int(0.25*w):int(0.75*w)] = 1
    inter = (cand_mask > 0) & (ref_mask > 0)
    union = (cand_mask > 0) | (ref_mask > 0)
    iou = inter.sum() / (union.sum() + 1e-5)
    return iou