Implement AutoAugment for Detection

[lezwon]

🚀 The feature

Implement Learning Data Augmentation Strategies for Object Detection
Refers to: #3817

Motivation, pitch

Good to have augmentation in Torchvision

Alternatives

No response

Additional context

No response

cc @vfdev-5 @datumbox

[lezwon]

@datumbox @vfdev-5 Based on our last conversation I suppose it would be best to start with the bbox_only_* augmentations for AA right? I could get started with implementing these in torchvision.prototype. Do let me know if that's alright.

1. Rotate_Only_BBoxes
2. ShearX_Only_BBoxes
3. ShearY_Only_BBoxes
4. TranslateX_Only_BBoxes
5. TranslateY_Only_BBoxes
6. Flip_Only_BBoxes
7. Solarize_Only_BBoxes
8. Equalize_Only_BBoxes
9. Cutout_Only_BBoxes

[datumbox]

@lezwon thanks a lot! It's worth coordinating with @vfdev-5 because he is literally landing changes at the same API.

Victor, how do you propose to do this? Would you prefer to do an implementation on References (similar to what we did with CopyPaste) and move it later on the new API?

[vfdev-5]

@datumbox @lezwon if we see AA detection code similar to classification one <=> single transformation class calling functional ops inside, then <op>_Only_BBoxes could be implemented as a special function that produces a list of image crops and we can apply <op>_image_tensor to them. I think we can go directly with new API.

By the way, what does Cutout_Only_BBoxes ? Erase data from target bbox ?

[lezwon]

@vfdev-5 Yep. Cutout does refer to erasing patches of data from the target box.

[ain-soph]

@vfdev-5
Here is a sample Cutout Implementation for image classification.
https://github.com/ain-soph/trojanzoo/blob/85730f629eddc17dcbc48218274b74c31daf5f99/trojanvision/utils/transform.py#L198-L226

torchvision doesn’t implement this because there is already a nearly equivalent RandomErasing. The only difference is for the boundary cases. And there is randomness of cutout area in RandomErasing. The class arguments are quite different as well.

[lezwon]

@vfdev-5 I'm sorry for the delay. I have been a bit occupied with other tasks and haven't been able to give any time to the implementation. I don't think I can pick this issue up anytime soon either. It would be best to assign someone else to it.

[ain-soph]

Any help needed? I’m interested to implement this.

[datumbox]

@ain-soph I think it would be awesome if you implement it.

Just checking again with @vfdev-5 and @pmeier that they don't have any concern.

[vfdev-5]

No problems from my side