search

keras-team / keras-cv

Industry-strength Computer Vision workflows with Keras
Other
1.01k stars 330 forks source link

Simple Copy-Paste Augmentation (Instance Segmentation) #28

Closed innat closed 1 year ago

innat commented 2 years ago

Demo: ds

Ref: https://github.com/tensorflow/tpu/tree/master/models/official/detection/projects/copy_paste

frostbyte012 commented 2 years ago

Can you explain a bit more I wanna know how this issue can be solved I want to contribute and I'm a first time contributor

bhack commented 2 years ago

Can you explain a bit more I wanna know how this issue can be solved I want to contribute and I'm a first time contributor

You need to read the related paper at https://arxiv.org/abs/2012.07177

innat commented 2 years ago

@frostbyte012 In addition to the above-mentioned paper, you can also look into pytorch implementation to get started.

all

MrinalTyagi commented 2 years ago

@innat @bhack Can I pick this up if no one else is working on it currently? Thank you

innat commented 2 years ago

@MrinalTyagi of course, anyone can. That would be great.

frostbyte012 commented 2 years ago

@innat @bhack is it possible to get the starting template / implementation in TensorFlow or Keras ? Any source or link?

innat commented 2 years ago

It's mentioned already ref. https://github.com/keras-team/keras-cv/issues/28#issue-1094848083

MrinalTyagi commented 2 years ago

@innat @bhack If I am not wrong, in this we will be adding a new data augmentation layer named CopyPasteAugmentation. In that, we will have initialization of parameters including sigma, blend, max objects that we can paste in an image, paste images, source images, x and y from which we have can apply copy-paste augmentation. This is the current scenario in my mind regarding it. Will update it accordingly during the course of time. Additionally, we should also provide a possible option to the users to save these augmented images as well as the information by how much the dataset was augmented as that can be really useful in some cases.

bhack commented 2 years ago

@MrinalTyagi use existing merged layers as a reference to grasp some code conventions.

@frostbyte012 Please coordinate with @MrinalTyagi you could collaborate on his branch if you are both interested.

MrinalTyagi commented 2 years ago

Got it. @frostbyte012 will share details of the draft pr with you soon with a basic template.

frostbyte012 commented 2 years ago

@MrinalTyagi I'm a newbie. Having a little difficulty in understanding it. If you can kindly help me together we can contribute and learn. @bhack I'd love to coordinate.

bhack commented 2 years ago

@LukeWood Can we temp. assign to them? So we could keep track on contribution-welcome assigned ticket vs the ones still looking for contribution.

LukeWood commented 2 years ago

@LukeWood Can we temp. assign to them? So we could keep track on contribution-welcome assigned ticket vs the ones still looking for contribution.

Sure thing

adhadse commented 2 years ago

@frostbyte012 Are you working on this issue? If not I would like to give this a try.

frostbyte012 commented 2 years ago

I'm not currently working on this, but it was taken by @MrinalTyagi I guess. I'm working on other issues. But I want to give a try as well.

adhadse commented 2 years ago

@frostbyte012 Sure, Give this a try. Mrinaltyagi has left this issue as said in his PR. Good Luck.

frostbyte012 commented 2 years ago

@adhadse Thanks!

tanzhenyu commented 1 year ago

@innat @bhack Can I pick this up if no one else is working on it currently? Thank you

I'm opening up this for contribution. Do you want to pick this up? It will be a little more complicated than other KPLs, given it requires some additional Tensor. But we already have copy paste augmentation in 3D KPL