Fix bugs during training runtime

configs/nus/hv_pointpillars_secfpn_sbn-all_4x8_20e_nus-3d.py → configs/nus/hv_pointpillars_secfpn_sbn-all_4x8_2x_nus-3d.py

@@ -130,7 +130,7 @@ input_modality = dict(
     use_lidar=True,
     use_depth=False,
     use_lidar_intensity=True,
-    use_camera=True,
+    use_camera=False,
 )
 db_sampler = dict(
     root_path=data_root,
@@ -156,23 +156,12 @@ train_pipeline = [
     dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
     dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
     dict(type='PointShuffle'),
-    dict(type='Normalize', **img_norm_cfg),
-    dict(type='Pad', size_divisor=32),
     dict(type='DefaultFormatBundle3D', class_names=class_names),
     dict(type='Collect3D', keys=['points', 'gt_bboxes_3d', 'gt_labels_3d']),
 ]
 test_pipeline = [
     dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
-    dict(
-        type='Resize',
-        img_scale=[
-            (1280, 720),
-        ],
-        multiscale_mode='value',
-        keep_ratio=True),
     dict(type='RandomFlip3D', flip_ratio=0),
-    dict(type='Normalize', **img_norm_cfg),
-    dict(type='Pad', size_divisor=32),
     dict(
         type='DefaultFormatBundle3D',
         class_names=class_names,
@@ -216,7 +205,7 @@ lr_config = dict(
     warmup='linear',
     warmup_iters=1000,
     warmup_ratio=1.0 / 1000,
-    step=[16, 19])
+    step=[20, 23])
 momentum_config = None
 checkpoint_config = dict(interval=1)
 # yapf:disable
@@ -229,10 +218,10 @@ log_config = dict(
     ])
 # yapf:enable
 # runtime settings
-total_epochs = 20
+total_epochs = 24
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
-work_dir = './work_dirs/pp_secfpn_80e'
+work_dir = './work_dirs/hv_pointpillars_secfpn_sbn-all_4x8_2x_nus-3d'
 load_from = None
 resume_from = None
 workflow = [('train', 1)]

mmdet3d/core/bbox/iou_calculators/iou3d_calculator.py

@@ -38,15 +38,16 @@ def bbox_overlaps_nearest_3d(bboxes1, bboxes2, mode='iou', is_aligned=False):
     """Calculate nearest 3D IoU
 
     Args:
-        bboxes1: Tensor, shape (N, 7) [x, y, z, h, w, l, ry]?
-        bboxes2: Tensor, shape (M, 7) [x, y, z, h, w, l, ry]?
+        bboxes1: Tensor, shape (N, 7+N) [x, y, z, h, w, l, ry, v]
+        bboxes2: Tensor, shape (M, 7+N) [x, y, z, h, w, l, ry, v]
         mode: mode (str): "iou" (intersection over union) or iof
             (intersection over foreground).
 
     Return:
         iou: (M, N) not support aligned mode currently
     """
-    assert bboxes1.size(-1) == bboxes2.size(-1) == 7
+    assert bboxes1.size(-1) >= 7
+    assert bboxes2.size(-1) >= 7
     column_index1 = bboxes1.new_tensor([0, 1, 3, 4, 6], dtype=torch.long)
     rbboxes1_bev = bboxes1.index_select(dim=-1, index=column_index1)
     rbboxes2_bev = bboxes2.index_select(dim=-1, index=column_index1)

mmdet3d/datasets/pipelines/train_aug.py

+import mmcv
 import numpy as np
 from mmcv.utils import build_from_cfg
 
@@ -34,7 +35,42 @@ class RandomFlip3D(RandomFlip):
         return gt_bboxes_3d, points
 
     def __call__(self, input_dict):
-        super(RandomFlip3D, self).__call__(input_dict)
+        # filp 2D image and its annotations
+        if 'flip' not in input_dict:
+            flip = True if np.random.rand() < self.flip_ratio else False
+            input_dict['flip'] = flip
+        if 'flip_direction' not in input_dict:
+            input_dict['flip_direction'] = self.direction
+        if input_dict['flip']:
+            # flip image
+            if 'img' in input_dict:
+                if isinstance(input_dict['img'], list):
+                    input_dict['img'] = [
+                        mmcv.imflip(
+                            img, direction=input_dict['flip_direction'])
+                        for img in input_dict['img']
+                    ]
+                else:
+                    input_dict['img'] = mmcv.imflip(
+                        input_dict['img'],
+                        direction=input_dict['flip_direction'])
+            # flip bboxes
+            for key in input_dict.get('bbox_fields', []):
+                input_dict[key] = self.bbox_flip(input_dict[key],
+                                                 input_dict['img_shape'],
+                                                 input_dict['flip_direction'])
+            # flip masks
+            for key in input_dict.get('mask_fields', []):
+                input_dict[key] = [
+                    mmcv.imflip(mask, direction=input_dict['flip_direction'])
+                    for mask in input_dict[key]
+                ]
+
+            # flip segs
+            for key in input_dict.get('seg_fields', []):
+                input_dict[key] = mmcv.imflip(
+                    input_dict[key], direction=input_dict['flip_direction'])
+
         if self.sync_2d:
             input_dict['pcd_flip'] = input_dict['flip']
         else:
@@ -50,6 +86,10 @@ class RandomFlip3D(RandomFlip):
             input_dict['points'] = points
         return input_dict
 
+    def __repr__(self):
+        return self.__class__.__name__ + '(flip_ratio={}, sync_2d={})'.format(
+            self.flip_ratio, self.sync_2d)
+
 
 @PIPELINES.register_module
 class ObjectSample(object):