Allow multiple RandomFlip operations in training pipeline

[SBCV]

Allow multiple RandomFlip operations in the training pipeline. The current implementation is limited to a single RandomFlip using a single direction.

Motivation

I've tried to add an additional RandomFlip step in the training pipeline, which results in something like this:

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=True),
    ...
    dict(type='RandomFlip', prob=0.5, direction="vertical"),
    dict(type='RandomFlip', prob=0.5, direction="horizontal"),
    ...
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]

However, in this case mmsegmentation does not apply any flips at all (neither vertical nor horizontal). If I remove one of the two lines, it works as expected.

I could provide a PR for this. However, in this case I need to know what kind of solution is preferable.

Possible approach 1:

dict(type='RandomFlip', prob=0.5, direction="vertical"),
dict(type='RandomFlip', prob=0.5, direction="horizontal"),

Possible approach 2: dict(type='RandomFlip', prob=(0.5, 0.5), direction=("vertical". "horizontal"))

[xiexinch]

Hi @SBCV, Thanks for your bug report, you might modify the following code to solve this problem simply. And if you have a better solution, welcome to create a PR. Personally, I think the second one is probably better :)

https://github.com/open-mmlab/mmsegmentation/blob/336435b80b26fb1273b5a4b5d597d1d65e2ff644/mmseg/datasets/pipelines/transforms.py#L361

if 'flip_direction' not in results or results['flip_direction'] != self.direction:
    results['flip_direction'] = self.direction

[SBCV]

Hi @xiexinch,

thank you for answer. I think the adjustment you have proposed does not properly address the corresponding meta data created by two consecutive flip operations. I'll push an appropriate PR tomorrow.

Bests

[MeowZheng]

I prefer approach 1, as the RandomFlip api is much easier to use. #1812 might increase complexity.

[SBCV]

@MeowZheng Added an alternative implementation according to your suggestion