Mappilary data prepration

[gurkirt]

Hi, Thank ​you for releasing the code. I am trying to reproduce your results on the Mapillary dataset. You provide script to convert 19 class cityscape into 16 class version in cvrn/init_citiscapes_19cls_to_16cls.py. Can this script also be used to preprocess mapillary dataset as well? What do I need to be able to evaluate your model to get the number shown in your paper? I see there is dataloader to mapillary dataset, but that doesn't seem to include conversion to 16 class dataset.

Can you explain the steps required?

Many thanks Gurkirt

[jxhuang0508]

Hi, we have uploaded the code that convert Mapillary labels into Cityscapes label space. Please find it in: CVRN/data loader and processing/init_vistas2cityscapes_format.py

Thanks!

[gurkirt]

Hi @jxhuang0508, thank you for providing the script. But I cannot find stuff_train.json or stuff_validation.json in the vista dataset. Did you generate those files yourself?

[jxhuang0508]

Hi, we have uploaded these two files onto Google Drive. You could generate them by yourself or download them via the Google Drive link: https://drive.google.com/file/d/1Fx9MHXQkIfKpKVDJ1B_drHx5ueocZfkG/view?usp=sharing https://drive.google.com/file/d/1e0-ru0QF1TEc9IlOJIqrUiw6bMDDEWYp/view?usp=sharing

[gurkirt]

Thank you @jxhuang0508, dataset preprocessing script seems to be working perfectly, still in the process though. However, I am unclear on the evaluation part now, I need a dataset loader like Cityscapes_CVRN for the vista dataset. Initially, I thought I could just Cityscapes_CVRN by changing the dataset path. But, it seems there are multiple hardcoded details only work for cityscape, e.g. gtFine, or paths.

Do I just need to fix those things or it will require more changes to run cross-dataset evaluation on evaluation?

[gurkirt]

I made the following changes vista_7cls.py to use it as dataloader. It seems to be indexing okay.

  -            'train': 'instancesonly_train_without_annotations.json',
  -            'val': 'instancesonly_gtFine_val.json',
  +            'train': 'instancesonly_train.json',
  +            'val': 'instancesonly_val.json',
  +            # 'train': 'instancesonly_train_without_annotations.json',
  +            # 'val': 'instancesonly_val.json.json',
               # 'test': 'image_info_test.json',
  
  -        self.panoptic_json_file = os.path.join(config.dataset.dataset_path, 'panoptic_val_16cls.json')
  -        self.panoptic_gt_folder = os.path.join(config.dataset.dataset_path, 'panoptic_val')
  +        self.panoptic_json_file = os.path.join(config.dataset.dataset_path, 'annotations/panoptic_val/panoptic_2018.json')
  +        self.panoptic_gt_folder = os.path.join(config.dataset.dataset_path, 'annotations/labels_val')
  
               for image_set, proposal_file in zip(image_sets, proposal_files):
                   dataset = JsonDataset('vistas_' + image_set,
                                         image_dir=self.image_dirs[image_set],
  -                                      anno_file=os.path.join(config.dataset.dataset_path, 'annotations_7cls', self.anno_files[image_set]))
  +                                      anno_file=os.path.join(config.dataset.dataset_path, 'annotations', self.anno_files[image_set]))
                   roidb = dataset.get_roidb(gt=True, proposal_file=proposal_file, crowd_filter_thresh=config.train.crowd_filter_thresh)
                   if flip:
                       if logger:
               self.dataset = JsonDataset('vistas_' + image_sets[0],
                                          image_dir=self.image_dirs[image_sets[0]],
  -                                       anno_file=os.path.join(config.dataset.dataset_path, 'annotations_7cls',
  +                                       anno_file=os.path.join(config.dataset.dataset_path, 'annotations',
                                          self.anno_files[image_sets[0]]))

However, I am concerned about

  -        self.panoptic_json_file = os.path.join(config.dataset.dataset_path, 'panoptic_val_16cls.json')
  -        self.panoptic_gt_folder = os.path.join(config.dataset.dataset_path, 'panoptic_val')
  +        self.panoptic_json_file = os.path.join(config.dataset.dataset_path, 'annotations/panoptic_val/panoptic_2018.json')
  +        self.panoptic_gt_folder = os.path.join(config.dataset.dataset_path, 'annotations/labels_val')

Is this okay?

[gurkirt]

@jxhuang0508, can you please provide panoptic_val_16cls.json?