[Feature] Support RMosaic

[heiyuxiaokai]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

1.Support RMosaic
(PS: Corresponding codes in mmdetection shows Mosaic must be used with RandomAffine. But I find Resize is also working. )

BC-breaking (Optional)

Does the modification introduce changes that break the back-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)

Use the train_pipeline as belowing:

train_pipeline = [
    dict(type='RMosaic', img_scale=(800, 800)),
    dict(type='RResize', img_scale=(800, 800)),
    dict(
        type='RRandomFlip',
        flip_ratio=[0.25, 0.25, 0.25],
        direction=['horizontal', 'vertical', 'diagonal'],
        version=angle_version),
    dict(
        type='PolyRandomRotate',
        rotate_ratio=0.5,
        angles_range=180,
        auto_bound=False,
        # rect_classes=[9, 11],
        version=angle_version),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

train_dataset = dict(
    _delete_ = True,
    type='MultiImageMixDataset',
    dataset=dict(
        type=dataset_type,
        classes=classes,
        ann_file=data_root + 'trainval/annfiles/',
        img_prefix=data_root + 'trainval/images/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='LoadAnnotations', with_bbox=True)
        ],
        version=angle_version
    ),
    pipeline=train_pipeline
)

data = dict(
    samples_per_gpu=4,
    workers_per_gpu=2,
    train=train_dataset,
    val=dict(
        classes=classes,
        ann_file=data_root + 'test/annfiles/',
        img_prefix=data_root + 'test/images/',
        pipeline=test_pipeline,
        version=angle_version),
    test=dict(
        classes=classes,
        ann_file=data_root + 'test/annfiles/',
        img_prefix=data_root + 'test/images/',
        pipeline=test_pipeline,
        version=angle_version))

Visualization

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. The documentation has been modified accordingly, like docstring or example tutorials.

[codecov]

Codecov Report

Merging #344 (e4239d7) into dev (d6ff82d) will decrease coverage by 0.12%.
The diff coverage is 17.24%.

@@            Coverage Diff             @@
##              dev     #344      +/-   ##
==========================================
- Coverage   30.86%   30.74%   -0.13%     
==========================================
  Files         119      119              
  Lines        7934     7989      +55     
  Branches     1220     1227       +7     
==========================================
+ Hits         2449     2456       +7     
- Misses       5384     5432      +48     
  Partials      101      101              

Flag	Coverage Δ
unittests	30.71% <17.24%> (-0.13%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmrotate/datasets/pipelines/transforms.py	65.77% <14.28%> (-17.17%)	⬇️
mmrotate/datasets/pipelines/__init__.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d6ff82d...e4239d7. Read the comment docs.

[zytx121]

Thank you for your PR! It's cool!
Could you please add the docstring for these functions?

[heiyuxiaokai]

Thank you for your PR! It's cool! Could you please add the docstring for these functions?

Ok, I'll update it later.

[liuyanyi]

Please follow https://github.com/open-mmlab/mmcv/blob/master/CONTRIBUTING.md to set pre-commit hook and fix the lint error.

[heiyuxiaokai]

Please follow https://github.com/open-mmlab/mmcv/blob/master/CONTRIBUTING.md to set pre-commit hook and fix the lint error.

Thanks! I'm confused about that.

[heiyuxiaokai]

Mosaic-transform outputs twice the input images. And resize-transform convert them to original sizes, which means all gt-boxes are half of inputs'. So I choose to set the min_box_size to 10, corresponding to min_box_size(5) of randomrotate.

From the view of visualizations, there seems no problem. But I got bad performance as：

Maybe mosaic should be used with randomcrop, and I'll try it.

[liuyanyi]

Mosaic-transform outputs twice the input images. And resize-transform convert them to original sizes, which means all gt-boxes are half of inputs'. So I choose to set the min_box_size to 10, corresponding to min_box_size(5) of randomrotate.

From the view of visualizations, there seems no problem. But I got bad performance as：

Maybe mosaic should be used with randomcrop, and I'll try it.

For mosaic and mixup, it should revert to orginal dataset in last few epoch, such as yolox in mmdet, you can have a try. I'm working on rrandom affine, it will release in few days.

[heiyuxiaokai]

Thanks for your explaining @liuyanyi . I'll have a try.

[zytx121]

LGTM

[nanhui69]

Thanks for your explaining @liuyanyi . I'll have a try.

does mosaic transform should combine with mixup? if that ,could you give a sample?