[transforms] Extends the list of supported data augmentations

[fg-mindee]

As discussed in #654, the artefact detection needs to improve its robustness. In order to do so and prevent overfitting, I would suggest gradually extending the list of our supported transformations:

• uniform photometric transformations (noise, blur, etc.) feat: Adds GaussianBlur, random font for CharGenerator and improves training scripts #758 feat: Added random noise augmentation to object detection #654 feat: Added ChannelShuffle transformation and fixes RandomCrop #768 feat: Added Gaussian Noise implementation in Tensorflow #771
• non-uniform photometric transformations (shadows, etc.) feat: Added shadow augmentation for all backends #811
• geometric transformations (flips, rotations, crops, etc.) feat: Added Random Horizontal Flip augmentation #773 [DRAFT] feat: Added RandomPerspective augmentation #799

Ideally, a given transform should be implemented in doctr/transforms/modules so that with the corresponding backend, we can do:

transfo = ....
pil_img = ... 
augmented_img = transfo(pil_img)

for transformations that only change the image.

and:

transfo = ....
pil_img = ... 
target = {...}
augmented_img, augmented_target = transfo(pil_img, target)

for transformations that alter the target. In doing so, that will work nicely with our Dataloaders 👍

[felixdittrich92]

@frgfm only @SiddhantBahuguna PR left or do you have other augmentations in mind ? :)

[frgfm]

For now, I think the random perspective would be the last one required yeah 👍

[felixdittrich92]

Hi @SiddhantBahuguna any way in near future that you finish your draft PR ? :)
Or should we keep it free to take ?

[SiddhantBahuguna]

Hi @SiddhantBahuguna any way in near future that you finish your draft PR ? :) Or should we keep it free to take ?

Greetings @felixdittrich92 , I am really sorry for the unwanted delay.
I will get it done this week positively :) Sorry again !

[felixdittrich92]

Sounds nice 🤗 👍

[felixdittrich92]

@SiddhantBahuguna Do you think you could finish your PR that we can close this ? :)

[felixdittrich92]

@odulcy-mindee I think we can close this ?
I don't see any issues with different perspectives (detection models) your dataset seems to capture this already well.
Or do you think we should add RandomPerspective as last one ftm ?

[odulcy-mindee]

Yeah, this issue can be closed, that's fine