[Feature] Support CopyPaste when mask is not available

[nijkah]

Motivation

closes #9686
There is a same feature in 2.x #8905

Some examples

Modification

Add new method to BaseBox

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repos?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

load_pipeline = [
    dict(type='LoadImageFromFile', backend_args=backend_args),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='RandomResize',
        scale=image_size,
        ratio_range=(0.8, 1.25),
        keep_ratio=True),
    dict(
        type='RandomCrop',
        crop_type='absolute_range',
        crop_size=image_size,
        recompute_bbox=True,
        allow_negative_crop=True),
    dict(type='FilterAnnotations', min_gt_bbox_wh=(1e-2, 1e-2)),
    dict(type='RandomFlip', prob=0.5),
    dict(type='Pad', size=image_size),
]
train_pipeline = [
    dict(type='CopyPaste', max_num_pasted=100, paste_by_box=True),
    dict(type='PackDetInputs')
]

train_dataloader = dict(
    dataset=dict(
        _delete_=True,
        type='MultiImageMixDataset',
        dataset=dict(
            type=dataset_type,
            data_root=data_root,
            ann_file='annotations/instances_train2017.json',
            data_prefix=dict(img='train2017/'),
            filter_cfg=dict(filter_empty_gt=True, min_size=32),
            pipeline=load_pipeline,
            backend_args=backend_args),
        pipeline=train_pipeline))

[jamiechoi1995]

Hi, after adapted your code, I got an error when running mmrotate 1.x

  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 396, in transform                                                       
    self._load_bboxes(results)                                                                                                                                           
  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 278, in _load_bboxes                                                    
    results['gt_bboxes'] = box_type_cls(gt_bboxes, dtype=torch.float32)                                                                                                  
TypeError: Can't instantiate abstract class QuadriBoxes with abstract methods create_masks 

[nijkah]

Hi, after adapted your code, I got an error when running mmrotate 1.x

  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 396, in transform                                                       
    self._load_bboxes(results)                                                                                                                                           
  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 278, in _load_bboxes                                                    
    results['gt_bboxes'] = box_type_cls(gt_bboxes, dtype=torch.float32)                                                                                                  
TypeError: Can't instantiate abstract class QuadriBoxes with abstract methods create_masks 

Hi, @jamiechoi1995. Because this transform depends on the box_type, every new box_type should define create_masks method. I will create a PR on MMRotate after merging this PR.

[jamiechoi1995]

Hi, after adapted your code, I got an error when running mmrotate 1.x

  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 396, in transform                                                       
    self._load_bboxes(results)                                                                                                                                           
  File "/home/jovyan/work/mmdetection-3.0.0/mmdet/datasets/transforms/loading.py", line 278, in _load_bboxes                                                    
    results['gt_bboxes'] = box_type_cls(gt_bboxes, dtype=torch.float32)                                                                                                  
TypeError: Can't instantiate abstract class QuadriBoxes with abstract methods create_masks 

Hi, @jamiechoi1995. Because this transform depends on the box_type, every new box_type should define create_masks method. I will create a PR on MMRotate after merging this PR.

Got it, thanks

[hhaAndroid]

@nijkah Please fix the CI

[nijkah]

@nijkah Please fix the CI

Hi, @hhaAndroid . I think the CI failed temporarily. I cannot see the error in the log.
Can I ask you to rerun it?

[nijkah]

Hi @hhaAndroid. I almost finished this PR. One thing left to consider.
What would be desirable when the config sets paste_by_box=True and results already contain gt_masks?

1. Throws warning or error message
2. Just use gt_masks as in the current code

[hhaAndroid]

Hi @hhaAndroid. I almost finished this PR. One thing left to consider. What would be desirable when the config sets paste_by_box=True and results already contain gt_masks?

1. Throws warning or error message
2. Just use gt_masks as in the current code

throws warning is ok and use gt_masks