[Fix] Fix bugs in RandomConcatDataset, SeqCropLikeStark and TridentSampling

mmtrack/datasets/builder.py

@@ -10,6 +10,7 @@
 from mmdet.datasets.samplers import (DistributedGroupSampler,
                                      DistributedSampler, GroupSampler)
 from torch.utils.data import DataLoader
+from torch.utils.data.dataset import ConcatDataset
 from torch.utils.data.sampler import RandomSampler
 
 from mmtrack.datasets.samplers.quota_sampler import DistributedQuotaSampler
@@ -56,10 +57,13 @@ def build_dataloader(dataset,
         DataLoader: A PyTorch dataloader.
     """
     rank, world_size = get_dist_info()
+    is_sotdataset = isinstance(dataset, BaseSOTDataset) or (
+        isinstance(dataset, ConcatDataset)
+        and isinstance(dataset.datasets[0], BaseSOTDataset))
     if dist:
         # ----- distributed train mode ------
         if shuffle:
-            if isinstance(dataset, BaseSOTDataset):
+            if is_sotdataset:
                 if samples_per_epoch is None:
                     sampler = DistributedSampler(
                         dataset, world_size, rank, shuffle=True)
@@ -90,7 +94,7 @@ def build_dataloader(dataset,
     else:
         # ----- non-distributed train mode ------
         if shuffle:
-            if isinstance(dataset, BaseSOTDataset):
+            if is_sotdataset:
                 if samples_per_epoch is None:
                     sampler = RandomSampler(dataset)
                 else:
@@ -104,10 +108,7 @@ def build_dataloader(dataset,
                 sampler = GroupSampler(dataset, samples_per_gpu)
         # ----- non-distributed test mode ------
         else:
-            if isinstance(dataset, BaseSOTDataset):
-                sampler = SOTVideoSampler(dataset)
-            else:
-                sampler = None
+            sampler = SOTVideoSampler(dataset) if is_sotdataset else None
 
         batch_size = num_gpus * samples_per_gpu
         num_workers = num_gpus * workers_per_gpu

mmtrack/datasets/dataset_wrappers.py

@@ -32,6 +32,8 @@ def __init__(self, dataset_cfgs, dataset_sampling_weights=None):
             ]
 
         datasets = [build_dataset(cfg) for cfg in dataset_cfgs]
+        self.CLASSES = datasets[0].CLASSES
+
         super().__init__(datasets)
 
     def __getitem__(self, ind):

mmtrack/datasets/pipelines/processing.py

@@ -117,14 +117,11 @@ def sampling_trident(self, video_visibility):
                     else:
                         min_ind, max_ind = search_ind - max_frame_range, \
                             search_ind
-                    # TODO: invisible objects are not used, because their
-                    # bboxes must cause interruption in the subsequent
-                    # pipeline.
                     extra_template_index = self.random_sample_inds(
                         video_visibility,
                         num_samples=1,
                         frame_range=[min_ind, max_ind],
-                        allow_invisible=True)[0]
+                        allow_invisible=False)[0]
 
                     extra_template_inds.append(extra_template_index)
 

mmtrack/datasets/pipelines/transforms.py

@@ -285,7 +285,7 @@ def __call__(self, results):
                 _results['img_shape'] = crop_img.shape
             _results['gt_bboxes'] = generated_bbox
             _results['seg_fields'] = ['padding_mask']
-            _results['pdding_mask'] = padding_mask
+            _results['padding_mask'] = padding_mask
             outs.append(_results)
         return outs