[RFC] Data Augmentation of object detection

[zhiqwang]

🚀 Feature

Hi, Recently I found a repo about data augmentation with excellent programming, named TensorPipe. But it's implemented with tensorflow, do you mind to add this similar features based in PyTorch in our quickvision repo?

Motivation

Augmentation play an important role in practical application especially when the collection of images and its labels is restricted by various reasons or the cost is very high.

Pitch

• Supports Mosiac Augmentation
• Supports CutOut

[oke-aditya]

Yes we can have such augmentation. Especially bounding box augmentations. As of now torchvision itself is getting upgrade for transforms. So maybe after a few releases we will have this.

We can plan for native ops module too.

[oke-aditya]

Once we have a nice nested tensor support in PyTorch as pointed out in torchvision (vmap, etc)
It wil be possible for torchvision itself to have params such as probablistic transforms, keep some batches from transforms.
This will make transforms even more powerful.

[zhiqwang]

That will be cool!