mmaction2
mmaction2 copied to clipboard
[PR Welcome] More data augmentation methods
Describe the feature More data augmentation methods.
Motivation When training our own models, we need to TRY EVERYTHING...
Additional context
- MMaction2 existing data augmentation methods:
- random flip.
- all kinds of crop and resize
- color jitter(brightness/contrast/saturation/hue)
Plans
- [x] imgaug #492
- [x] mixup and cutmix #681
- [x] RandAugment by imgaug #585
- [x] PytorchVideo Transforms. #1008
- [ ] albumentations
- [ ] videomix
More image augmentation methods could refer to PaddleClas.
Thanks for the enthusiastic! Here is a practical way:
- Make a list of candidate augmentations. For each augmentation, find its original paper, together with a reference implementation.
- Make individual pull requests for each one of the list.
You may use this issue to track progress.
Yeah, Thanks for your proposal. It is better to provide a list of candidate augmentations so that we can also help to implement some of these ! @irvingzhang0512
this paper shows that random rotation may help
In image classification, rotation works to some extent. But color jittering sometimes does not work.
Whether they work or not, it does not affect whether we implement them.
edit: @dreamerlin had some code on color jittering
Supporting rotation would be great.
I also wanted to have some camera transformations, such as
- push in
- pull out
- pan
- tilt
Do you have interest in implementing them?
Supporting rotation would be great.
I also wanted to have some camera transformations, such as
- push in
- pull out
- pan
- tilt
Do you have interest in implementing them?
@innerlee I'm a little busy until Jan. 20th. If no one implemente these then, I'll have a try. Before that, I'll implement tsm-mobilenet and support imgaug/albumentation in pipeline.
No hurry.
There is also lots of movie transition fx that can do some fancy temporal "mixup"
Find an interesting paper here. This paper proposes a video augmentation strategy called VideoMix
.
haha
PytorchVideo transform random_resized_crop support shift
mode, which looks like the movie transforms. Maybe we can support this.
@innerlee @dreamerlin