In sim2real task, the source and target are not describing the same scene which is meaningless for transfer learning

[drowning-in-codes]

Hi, thank you for your excellent work. I noticed that source and target data were shuffled in the domain adaptation training.
The simulated opv2v and the real-world dataset v2v4 do not just have domain differences. For example, in the same frame, in opv2v, it could be three CAVs on a curvy road, while in v2v4, there could be one CAV on a straight road, so they have different labels.

 source_train_loader = DataLoader(source_opencood_train_dataset,
                                     batch_size=hypes['train_params'][
                                         'batch_size'],
                                     num_workers=8,
                                     collate_fn=source_opencood_train_dataset.collate_batch_train,
                                     shuffle=True,
                                     pin_memory=False,
                                     drop_last=True)
  target_train_loader = DataLoader(target_opencood_train_dataset,
                                     batch_size=hypes['train_params'][
                                         'batch_size'],
                                     num_workers=8,
                                     collate_fn=source_opencood_train_dataset.collate_batch_train,
                                     shuffle=True,
                                     pin_memory=False,
                                     drop_last=True)

The domain adaptation configuration YAML file is as follows.

name: corpbevtlidar_da
root_dir: '/data/opv2v/rain' # the path of the source domain
root_dir_target: '/data/v2v4real/train' # the path of the target domain
validate_dir: '/data/v2v4real/val'

I think it would make sense to combine the source data and target data into one scene with different point cloud distributions.
Appreciate again.