[Refactor] update data flow and ut (open-mmlab#1776)

* update data flow and ut

* update ut

* update code

* fix mapping bug

* fix comments

configs/_base_/datasets/kitti-mono3d.py

@@ -26,8 +26,8 @@
     dict(
         type='Pack3DDetInputs',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_3d', 'gt_labels_3d',
-            'centers_2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'gt_bboxes_3d',
+            'gt_labels_3d', 'centers_2d', 'depths'
         ]),
 ]
 test_pipeline = [

configs/_base_/datasets/nus-mono3d.py

@@ -37,8 +37,8 @@
     dict(
         type='Pack3DDetInputs',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'attr_labels', 'gt_bboxes_3d',
-            'gt_labels_3d', 'centers_2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'attr_labels',
+            'gt_bboxes_3d', 'gt_labels_3d', 'centers_2d', 'depths'
         ]),
 ]
 

configs/fcos3d/fcos3d_r101-caffe-dcn_fpn_head-gn_8xb2-1x_nus-mono3d.py

@@ -42,8 +42,8 @@
     dict(
         type='Pack3DDetInputs',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'attr_labels', 'gt_bboxes_3d',
-            'gt_labels_3d', 'centers_2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'attr_labels',
+            'gt_bboxes_3d', 'gt_labels_3d', 'centers_2d', 'depths'
         ]),
 ]
 test_pipeline = [

configs/pgd/pgd_r101-caffe_fpn_head-gn_16xb2-1x_nus-mono3d.py

@@ -63,8 +63,8 @@
     dict(
         type='Collect3D',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'attr_labels', 'gt_bboxes_3d',
-            'gt_labels_3d', 'centers2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'attr_labels',
+            'gt_bboxes_3d', 'gt_labels_3d', 'centers2d', 'depths'
         ]),
 ]
 test_pipeline = [

configs/pgd/pgd_r101-caffe_fpn_head-gn_4xb3-4x_kitti-mono3d.py

@@ -96,8 +96,8 @@
     dict(
         type='Pack3DDetInputs',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_3d', 'gt_labels_3d',
-            'centers_2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'gt_bboxes_3d',
+            'gt_labels_3d', 'centers_2d', 'depths'
         ]),
 ]
 test_pipeline = [

configs/smoke/smoke_dla34_dlaneck_gn-all_4xb8-6x_kitti-mono3d.py

@@ -32,8 +32,8 @@
     dict(
         type='Pack3DDetInputs',
         keys=[
-            'img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_3d', 'gt_labels_3d',
-            'centers_2d', 'depths'
+            'img', 'gt_bboxes', 'gt_bboxes_labels', 'gt_bboxes_3d',
+            'gt_labels_3d', 'centers_2d', 'depths'
         ]),
 ]
 test_pipeline = [

mmdet3d/apis/inference.py

@@ -250,7 +250,7 @@ def inference_multi_modality_detector(model: nn.Module,
         results = model.test_step(data)
 
     for index in range(len(data)):
-        meta_info = data[index]['data_sample'].metainfo
+        meta_info = data[index]['data_samples'].metainfo
         results[index].set_metainfo(meta_info)
 
     if not is_batch:
@@ -320,7 +320,7 @@ def inference_mono_3d_detector(model: nn.Module,
         results = model.test_step(data)
 
     for index in range(len(data)):
-        meta_info = data[index]['data_sample'].metainfo
+        meta_info = data[index]['data_samples'].metainfo
         results[index].set_metainfo(meta_info)
 
     if not is_batch:

mmdet3d/datasets/det3d_dataset.py

@@ -187,12 +187,13 @@ def parse_ann_info(self, info: dict) -> Optional[dict]:
         # in `transforms`
         name_mapping = {
             'bbox_label_3d': 'gt_labels_3d',
+            'bbox_label': 'gt_bboxes_labels',
+            'bbox': 'gt_bboxes',
             'bbox_3d': 'gt_bboxes_3d',
             'depth': 'depths',
             'center_2d': 'centers_2d',
             'attr_label': 'attr_labels'
         }
-
         instances = info['instances']
         # empty gt
         if len(instances) == 0:
@@ -203,13 +204,18 @@ def parse_ann_info(self, info: dict) -> Optional[dict]:
             for ann_name in keys:
                 temp_anns = [item[ann_name] for item in instances]
                 # map the original dataset label to training label
-                if 'label' in ann_name:
+                if 'label' in ann_name and ann_name != 'attr_label':
                     temp_anns = [
                         self.label_mapping[item] for item in temp_anns
                     ]
                 if ann_name in name_mapping:
                     ann_name = name_mapping[ann_name]
-                temp_anns = np.array(temp_anns)
+
+                if 'label' in ann_name:
+                    temp_anns = np.array(temp_anns).astype(np.int64)
+                else:
+                    temp_anns = np.array(temp_anns).astype(np.float32)
+
                 ann_info[ann_name] = temp_anns
             ann_info['instances'] = info['instances']
         return ann_info
@@ -310,7 +316,7 @@ def prepare_data(self, index):
             # after pipeline drop the example with empty annotations
             # return None to random another in `__getitem__`
             if example is None or len(
-                    example['data_sample'].gt_instances_3d.labels_3d) == 0:
+                    example['data_samples'].gt_instances_3d.labels_3d) == 0:
                 return None
         return example
 

mmdet3d/datasets/transforms/formating.py

@@ -97,8 +97,8 @@ def transform(self, results: Union[dict,
                 - points
                 - img
 
-            - 'data_sample' (obj:`Det3DDataSample`): The annotation info of the
-              sample.
+            - 'data_samples' (obj:`Det3DDataSample`): The annotation info of
+                the sample.
         """
         # augtest
         if isinstance(results, list):
@@ -131,8 +131,8 @@ def pack_single_results(self, results):
                 - points
                 - img
 
-            - 'data_sample' (obj:`Det3DDataSample`): The annotation info of the
-              sample.
+            - 'data_samples' (obj:`Det3DDataSample`): The annotation info
+              of the sample.
         """
         # Format 3D data
         if 'points' in results:
@@ -213,7 +213,7 @@ def pack_single_results(self, results):
             data_sample.eval_ann_info = None
 
         packed_results = dict()
-        packed_results['data_sample'] = data_sample
+        packed_results['data_samples'] = data_sample
         packed_results['inputs'] = inputs
 
         return packed_results

mmdet3d/datasets/transforms/loading.py

@@ -746,17 +746,8 @@ def _load_bboxes(self, results: dict) -> None:
         Returns:
             dict: The dict contains loaded bounding box annotations.
         """
-        gt_bboxes = []
-        for instance in results['instances']:
-            gt_bboxes.append(instance['bbox'])
-        if len(gt_bboxes) == 0:
-            results['gt_bboxes'] = np.zeros((0, 4), dtype=np.float32)
-        else:
-            results['gt_bboxes'] = np.array(
-                gt_bboxes, dtype=np.float32).reshape((-1, 4))
 
-        if 'eval_ann_info' in results:
-            results['eval_ann_info']['gt_bboxes'] = results['gt_bboxes']
+        results['gt_bboxes'] = results['ann_info']['gt_bboxes']
 
     def _load_labels(self, results: dict) -> None:
         """Private function to load label annotations.
@@ -767,17 +758,7 @@ def _load_labels(self, results: dict) -> None:
         Returns:
             dict: The dict contains loaded label annotations.
         """
-        gt_bboxes_labels = []
-        for instance in results['instances']:
-            gt_bboxes_labels.append(instance['bbox_label'])
-        if len(gt_bboxes_labels) == 0:
-            results['gt_bboxes_labels'] = np.zeros((0, ), dtype=np.int64)
-        else:
-            results['gt_bboxes_labels'] = np.array(
-                gt_bboxes_labels, dtype=np.int64)
-        if 'eval_ann_info' in results:
-            results['eval_ann_info']['gt_bboxes_labels'] = results[
-                'gt_bboxes_labels']
+        results['gt_bboxes_labels'] = results['ann_info']['gt_bboxes_labels']
 
     def transform(self, results: dict) -> dict:
         """Function to load multiple types annotations.

mmdet3d/evaluation/metrics/indoor_metric.py

@@ -37,32 +37,28 @@ def __init__(self,
             prefix=prefix, collect_device=collect_device)
         self.iou_thr = iou_thr
 
-    def process(self, data_batch: Sequence[dict],
-                predictions: Sequence[dict]) -> None:
+    def process(self, data_batch: dict, data_samples: Sequence[dict]) -> None:
         """Process one batch of data samples and predictions.
 
         The processed results should be stored in ``self.results``,
         which will be used to compute the metrics when all batches
         have been processed.
 
         Args:
-            data_batch (Sequence[dict]): A batch of data
-                from the dataloader.
-            predictions (Sequence[dict]): A batch of outputs from
+            data_batch (dict): A batch of data from the dataloader.
+            data_samples (Sequence[dict]): A batch of outputs from
                 the model.
         """
-        batch_eval_anns = [
-            item['data_sample']['eval_ann_info'] for item in data_batch
-        ]
-        for eval_ann, pred_dict in zip(batch_eval_anns, predictions):
-            pred_3d = pred_dict['pred_instances_3d']
+        for data_sample in data_samples:
+            pred_3d = data_sample['pred_instances_3d']
+            eval_ann_info = data_sample['eval_ann_info']
             cpu_pred_3d = dict()
             for k, v in pred_3d.items():
                 if hasattr(v, 'to'):
                     cpu_pred_3d[k] = v.to('cpu')
                 else:
                     cpu_pred_3d[k] = v
-            self.results.append((eval_ann, cpu_pred_3d))
+            self.results.append((eval_ann_info, cpu_pred_3d))
 
     def compute_metrics(self, results: list) -> Dict[str, float]:
         """Compute the metrics from processed results.
@@ -121,28 +117,24 @@ def __init__(self,
             prefix=prefix, collect_device=collect_device)
         self.iou_thr = iou_thr
 
-    def process(self, data_batch: Sequence[dict],
-                predictions: Sequence[dict]) -> None:
+    def process(self, data_batch: dict, data_samples: Sequence[dict]) -> None:
         """Process one batch of data samples and predictions.
 
         The processed results should be stored in ``self.results``,
         which will be used to compute the metrics when all batches
         have been processed.
 
         Args:
-            data_batch (Sequence[dict]): A batch of data
-                from the dataloader.
+            data_batch (dict): A batch of data from the dataloader.
             predictions (Sequence[dict]): A batch of outputs from
                 the model.
         """
-        batch_eval_anns = [
-            item['data_sample']['eval_ann_info'] for item in data_batch
-        ]
-        for eval_ann, pred_dict in zip(batch_eval_anns, predictions):
-            pred = pred_dict['pred_instances']
+        for data_sample in data_samples:
+            pred = data_sample['pred_instances']
+            eval_ann_info = data_sample['eval_ann_info']
             ann = dict(
-                labels=eval_ann['gt_bboxes_labels'],
-                bboxes=eval_ann['gt_bboxes'])
+                labels=eval_ann_info['gt_bboxes_labels'],
+                bboxes=eval_ann_info['gt_bboxes'])
 
             pred_bboxes = pred['bboxes'].cpu().numpy()
             pred_scores = pred['scores'].cpu().numpy()

mmdet3d/evaluation/metrics/instance_seg_metric.py

@@ -29,32 +29,28 @@ def __init__(self,
         super(InstanceSegMetric, self).__init__(
             prefix=prefix, collect_device=collect_device)
 
-    def process(self, data_batch: Sequence[dict],
-                predictions: Sequence[dict]) -> None:
+    def process(self, data_batch: dict, data_samples: Sequence[dict]) -> None:
         """Process one batch of data samples and predictions.
 
         The processed results should be stored in ``self.results``,
         which will be used to compute the metrics when all batches
         have been processed.
 
         Args:
-            data_batch (Sequence[dict]): A batch of data
-                from the dataloader.
-            predictions (Sequence[dict]): A batch of outputs from
+            data_batch (dict): A batch of data from the dataloader.
+            data_samples (Sequence[dict]): A batch of outputs from
                 the model.
         """
-        batch_eval_anns = [
-            item['data_sample']['eval_ann_info'] for item in data_batch
-        ]
-        for eval_ann, pred_dict in zip(batch_eval_anns, predictions):
-            pred_3d = pred_dict['pred_pts_seg']
+        for data_sample in data_samples:
+            pred_3d = data_sample['pred_pts_seg']
+            eval_ann_info = data_sample['eval_ann_info']
             cpu_pred_3d = dict()
             for k, v in pred_3d.items():
                 if hasattr(v, 'to'):
                     cpu_pred_3d[k] = v.to('cpu')
                 else:
                     cpu_pred_3d[k] = v
-            self.results.append((eval_ann, cpu_pred_3d))
+            self.results.append((eval_ann_info, cpu_pred_3d))
 
     def compute_metrics(self, results: list) -> Dict[str, float]:
         """Compute the metrics from processed results.

mmdet3d/evaluation/metrics/kitti_metric.py

@@ -137,29 +137,30 @@ def convert_annos_to_kitti_annos(
             data_annos[i]['kitti_annos'] = kitti_annos
         return data_annos
 
-    def process(self, data_batch: Sequence[dict],
-                predictions: Sequence[dict]) -> None:
+    def process(self, data_batch: dict, data_samples: Sequence[dict]) -> None:
         """Process one batch of data samples and predictions.
 
         The processed results should be stored in ``self.results``,
         which will be used to compute the metrics when all batches
         have been processed.
 
         Args:
-            data_batch (Sequence[dict]): A batch of data
-                from the dataloader.
-            predictions (Sequence[dict]): A batch of outputs from
+            data_batch (dict): A batch of data from the dataloader.
+            data_samples (Sequence[dict]): A batch of outputs from
                 the model.
         """
-        assert len(data_batch) == len(predictions)
-        for data, pred in zip(data_batch, predictions):
+
+        for data_sample in data_samples:
             result = dict()
-            for pred_result in pred:
-                for attr_name in pred[pred_result]:
-                    pred[pred_result][attr_name] = pred[pred_result][
-                        attr_name].to(self.collect_device)
-                result[pred_result] = pred[pred_result]
-            sample_idx = data['data_sample']['sample_idx']
+            pred_3d = data_sample['pred_instances_3d']
+            pred_2d = data_sample['pred_instances']
+            for attr_name in pred_3d:
+                pred_3d[attr_name] = pred_3d[attr_name].to('cpu')
+            result['pred_instances_3d'] = pred_3d
+            for attr_name in pred_2d:
+                pred_2d[attr_name] = pred_2d[attr_name].to('cpu')
+            result['pred_instances'] = pred_2d
+            sample_idx = data_sample['sample_idx']
             result['sample_idx'] = sample_idx
         self.results.append(result)