dataset read aligned bbox

configs/_base_/datasets/scannet-3d-18class.py

@@ -16,18 +16,16 @@
  use_dim=[0, 1, 2]),
  dict(
  type='LoadAnnotations3D',
- with_bbox_3d=False,
- with_label_3d=False,
+ with_bbox_3d=True,
+ with_label_3d=True,
  with_mask_3d=True,
  with_seg_3d=True),
+ dict(type='GlobalAlignment', rotation_axis=2),
  dict(
  type='PointSegClassMapping',
  valid_cat_ids=(3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16, 24, 28, 33, 34,
  36, 39),
  max_cat_id=40),
- dict(
- type='GlobalAlignment', rotation_axis=2,
- ignore_index=len(class_names)),
  dict(type='IndoorPointSample', num_points=40000),
  dict(
  type='RandomFlip3D',
@@ -54,9 +52,7 @@
  shift_height=True,
  load_dim=6,
  use_dim=[0, 1, 2]),
- dict(
- type='GlobalAlignment', rotation_axis=2,
- ignore_index=len(class_names)),
+ dict(type='GlobalAlignment', rotation_axis=2),
  dict(
  type='MultiScaleFlipAug3D',
  img_scale=(1333, 800),
@@ -93,16 +89,11 @@
  use_dim=[0, 1, 2]),
  dict(
  type='LoadAnnotations3D',
- with_bbox_3d=False,
- with_label_3d=False,
- with_mask_3d=True,
- with_seg_3d=True),
- dict(type='PointSegClassMapping', valid_cat_ids=valid_class_ids),
- dict(
- type='GlobalAlignment',
- rotation_axis=2,
- ignore_index=len(class_names),
- extract_bbox=True),
+ with_bbox_3d=True,
+ with_label_3d=True,
+ with_mask_3d=False,
+ with_seg_3d=False),
+ dict(type='GlobalAlignment', rotation_axis=2),
  dict(
  type='DefaultFormatBundle3D',
  class_names=class_names,

mmdet3d/datasets/pipelines/transforms_3d.py

@@ -296,28 +296,20 @@ def __repr__(self):
 @PIPELINES.register_module()
 class GlobalAlignment(object):
  """Apply global alignment to 3D scene points by rotation and translation.
- Extract 3D bboxes from the aligned points and instance mask if provided.
 
  Args:
  rotation_axis (int): Rotation axis for points and bboxes rotation.
- ignore_index (int): Label index for which we won't extract bboxes.
- extract_bbox (bool): Whether extract new ground-truth bboxes after \
- alignment. This requires instance and semantic mask inputs.
- Defaults to False.
 
  Note:
- This function should be called after PointSegClassMapping in pipeline.
  We do not record the applied rotation and translation as in \
  GlobalRotScaleTrans. Because usually, we do not need to reverse \
  the alignment step.
  For example, ScanNet 3D detection task uses aligned ground-truth \
  bounding boxes for evaluation.
  """
 
- def __init__(self, rotation_axis, ignore_index, extract_bbox=False):
+ def __init__(self, rotation_axis):
  self.rotation_axis = rotation_axis
- self.ignore_index = ignore_index
- self.extract_bbox = extract_bbox
 
  def _trans_points(self, input_dict, trans_factor):
  """Private function to translate points.
@@ -357,74 +349,6 @@ def _check_rot_mat(self, rot_mat):
  is_valid &= (rot_mat[:, self.rotation_axis] == valid_array).all()
  assert is_valid, f'invalid rotation matrix {rot_mat}'
 
- def _bbox_from_points(self, points):
- """Get the bounding box of a set of points.
-
- Args:
- points (np.ndarray): A set of points belonging to one instance.
-
- Returns:
- np.ndarray: A bounding box of input points. We use origin as \
- (0.5, 0.5, 0.5) without yaw.
- """
- xmin = np.min(points[:, 0])
- ymin = np.min(points[:, 1])
- zmin = np.min(points[:, 2])
- xmax = np.max(points[:, 0])
- ymax = np.max(points[:, 1])
- zmax = np.max(points[:, 2])
- bbox = np.array([(xmin + xmax) / 2, (ymin + ymax) / 2,
- (zmin + zmax) / 2, xmax - xmin, ymax - ymin,
- zmax - zmin])
- return bbox
-
- def _extract_bboxes(self, input_dict):
- """Extract bounding boxes from points, semantic mask and instance mask.
-
- Args:
- input_dict (dict): Result dict from loading pipeline.
-
- Returns:
- dict: Results after extracting bboxes, keys in \
- input_dict['bbox3d_fields'] are updated in the dict.
- """
- # TODO: this function is only used in ScanNet-Det pipeline currently
- # TODO: we only extract gt_bboxes_3d which is DepthInstance3DBoxes
- from mmdet3d.core.bbox import DepthInstance3DBoxes
-
- assert 'pts_instance_mask' in input_dict.keys(), \
- 'instance mask is not provided in GlobalAlignment'
- assert 'pts_semantic_mask' in input_dict.keys(), \
- 'semantic mask is not provided in GlobalAlignment'
-
- coords = input_dict['points'].coord.numpy()
- inst_mask = input_dict['pts_instance_mask']
- sem_mask = input_dict['pts_semantic_mask']
-
- # select points from valid categories where we want to extract bboxes
- valid_cat_mask = (sem_mask != self.ignore_index)
- inst_ids = np.unique(inst_mask[valid_cat_mask]) # ids of valid insts
- instance_bboxes = np.zeros((inst_ids.shape[0], 7))
- inst_id2cat_id = {
- inst_id: sem_mask[inst_mask == inst_id][0]
- for inst_id in inst_ids
- }
- for bbox_idx, inst_id in enumerate(inst_ids):
- cat_id = inst_id2cat_id[inst_id]
- inst_coords = coords[inst_mask == inst_id]
- bbox = self._bbox_from_points(inst_coords)
- instance_bboxes[bbox_idx, :6] = bbox
- instance_bboxes[bbox_idx, 6] = cat_id
-
- if 'gt_bboxes_3d' not in input_dict['bbox3d_fields']:
- input_dict['bbox3d_fields'].append('gt_bboxes_3d')
- input_dict['gt_bboxes_3d'] = DepthInstance3DBoxes(
- instance_bboxes[:, :6],
- box_dim=6,
- with_yaw=False,
- origin=(0.5, 0.5, 0.5))
- input_dict['gt_labels_3d'] = instance_bboxes[:, 6].astype(np.long)
-
  def __call__(self, input_dict):
  """Call function to shuffle points.
 
@@ -447,16 +371,12 @@ def __call__(self, input_dict):
  self._check_rot_mat(rot_mat)
  self._rot_points(input_dict, rot_mat)
  self._trans_points(input_dict, trans_vec)
- if self.extract_bbox:
- self._extract_bboxes(input_dict)
 
  return input_dict
 
  def __repr__(self):
  repr_str = self.__class__.__name__
- repr_str += f'(rotation_axis={self.rotation_axis},'
- repr_str += f' ignore_index={self.ignore_index},'
- repr_str += f' extract_bbox={self.extract_bbox})'
+ repr_str += f'(rotation_axis={self.rotation_axis})'
  return repr_str
 
 

mmdet3d/datasets/scannet_dataset.py

@@ -154,89 +154,6 @@ def _get_axis_align_matrix(info):
  'use new pre-process scripts to re-generate ScanNet data')
  return np.eye(4).astype(np.float32)
 
- def evaluate(self,
- results,
- metric=None,
- iou_thr=(0.25, 0.5),
- logger=None,
- show=False,
- out_dir=None,
- pipeline=None):
- """Evaluate.
-
- Evaluation in indoor protocol.
- Since ScanNet detection data pipeline re-computes ground-truth boxes,
- we can't directly use gt_bboxes from self.data_infos.
-
- Args:
- results (list[dict]): List of results.
- metric (str | list[str]): Metrics to be evaluated.
- iou_thr (list[float]): AP IoU thresholds.
- show (bool): Whether to visualize.
- Default: False.
- out_dir (str): Path to save the visualization results.
- Default: None.
- pipeline (list[dict], optional): raw data loading for showing.
- Default: None.
-
- Returns:
- dict: Evaluation results.
- """
- from mmdet3d.core.evaluation import indoor_eval
- assert isinstance(
- results, list), f'Expect results to be list, got {type(results)}.'
- assert len(results) > 0, 'Expect length of results > 0.'
- assert len(results) == len(self.data_infos)
- assert isinstance(
- results[0], dict
- ), f'Expect elements in results to be dict, got {type(results[0])}.'
- # load gt_bboxes via pipeline
- pipeline = self._get_pipeline(pipeline)
- gt_bboxes = [
- self._extract_data(
- i, pipeline, ['gt_bboxes_3d', 'gt_labels_3d'], load_annos=True)
- for i in range(len(self.data_infos))
- ]
- gt_annos = [self._build_annos(*gt_bbox) for gt_bbox in gt_bboxes]
- label2cat = {i: cat_id for i, cat_id in enumerate(self.CLASSES)}
- ret_dict = indoor_eval(
- gt_annos,
- results,
- iou_thr,
- label2cat,
- logger=logger,
- box_type_3d=self.box_type_3d,
- box_mode_3d=self.box_mode_3d)
- if show:
- self.show(results, out_dir, pipeline=pipeline)
-
- return ret_dict
-
- @staticmethod
- def _build_annos(gt_bboxes, gt_labels):
- """Transform gt bboxes and labels into self.data_infos['annos'] format.
-
- Args:
- gt_bboxes (:obj:`BaseInstance3DBoxes`): \
- 3D bounding boxes in Depth coordinate
- gt_labels (torch.Tensor): Labels of boxes.
-
- Returns:
- dict: annotations including the following keys
-
- - gt_boxes_upright_depth (np.ndarray): 3D bounding boxes.
- - class (np.ndarray): Labels of boxes.
- - gt_num (int): Number of boxes.
- """
- bbox = gt_bboxes.tensor.numpy()[:, :6].copy() # drop yaw dimension
- bbox[..., 2] += bbox[..., 5] / 2 # bottom center to gravity center
- anno = {
- 'gt_boxes_upright_depth': bbox,
- 'class': gt_labels.numpy(),
- 'gt_num': gt_labels.shape[0]
- }
- return anno
-
  def _build_default_pipeline(self):
  """Build the default pipeline for this dataset."""
  pipeline = [
@@ -248,19 +165,11 @@ def _build_default_pipeline(self):
  use_dim=[0, 1, 2]),
  dict(
  type='LoadAnnotations3D',
- with_bbox_3d=False,
- with_label_3d=False,
- with_mask_3d=True,
- with_seg_3d=True),
- dict(
- type='PointSegClassMapping',
- valid_cat_ids=(3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16, 24, 28,
- 33, 34, 36, 39)),
- dict(
- type='GlobalAlignment',
- rotation_axis=2,
- ignore_index=len(self.CLASSES),
- extract_bbox=True),
+ with_bbox_3d=True,
+ with_label_3d=True,
+ with_mask_3d=False,
+ with_seg_3d=False),
+ dict(type='GlobalAlignment', rotation_axis=2),
  dict(
  type='DefaultFormatBundle3D',
  class_names=self.CLASSES,
@@ -287,10 +196,9 @@ def show(self, results, out_dir, show=True, pipeline=None):
  data_info = self.data_infos[i]
  pts_path = data_info['pts_path']
  file_name = osp.split(pts_path)[-1].split('.')[0]
- points, gt_bboxes = self._extract_data(
- i, pipeline, ['points', 'gt_bboxes_3d'], load_annos=True)
- points = points.numpy()
- gt_bboxes = gt_bboxes.tensor.numpy()
+ points = self._extract_data(
+ i, pipeline, 'points', load_annos=True).numpy()
+ gt_bboxes = self.get_ann_info(i)['gt_bboxes_3d'].tensor.numpy()
  pred_bboxes = result['boxes_3d'].tensor.numpy()
  show_result(points, gt_bboxes, pred_bboxes, out_dir, file_name,
  show)