CutOut should support fill_mode="mirror"

[LukeWood]

This is actually a great "research" project in that this should strictly be better than using noise and black pixels.

[innat]

Can you add some references, sounds interesting.

[LukeWood]

There are references out there for mirroring in random translation and rotation, but not random erasing or cut out. The logic should in theory extend to cutout, but I do not have definitive proof. That's why this is also somewhat of a research effort. Happy to collaborate with anyone on benchmarks if they are interested...

[bhack]

Are you talking about (https://arxiv.org/pdf/1708.04552v2.pdf):

or something else?

Edit: Ok sorry it was in the title but not in the description. It is the "fill_mode"

[bhack]

We have some experimental results with patch gaussian as an alternative "fill_mode" for cutout:

https://arxiv.org/abs/1906.02611

[innat]

The fill_mode="mirror" sounds like a fill_mode="reflect".

[bhack]

The patch gaussian has a good citations threshod (>100).

There is a TF implementation at page 15.

I don't know if It Is better to spent time on something that has already some experimental results.

[bhack]

https://github.com/tensorflow/models/blob/master/research/object_detection/core/preprocessor.py#L2793

[innat]

@bhack Thanks for sharing this information, very useful. It's definitely an effective option.

Cutout, which improves clean accuracy but not robustness, and additive Gaussian noise, which improves robustness but hurts accuracy. To overcome this trade-off, we introduce Patch Gaussian, a simple augmentation scheme that adds noise to randomly selected patches in an input image.

[github-actions[bot]]

This issue is stale because it has been open for 180 days with no activity. It will be closed if no further activity occurs. Thank you.

[sachinprasadhs]

Thanks for reporting the issue! We have consolidated the development of KerasCV into the new KerasHub package, which supports image, text, and multi-modal models. Please read https://github.com/keras-team/keras-hub/issues/1831. KerasHub will support all the core functionality of KerasCV.

KerasHub can be installed with !pip install -U keras-hub. Documentation and guides are available at keras.io/keras_hub.

With our focus shifted to KerasHub, we are not planning any further development or releases in KerasCV. If you encounter a KerasCV feature that is missing from KerasHub, or would like to propose an addition to the library, please file an issue with KerasHub.

[innat]

@sachinprasadhs

If you encounter a KerasCV feature that is missing from KerasHub, or would like to propose an addition to the library, please file an issue with KerasHub.

Firstly, identifying which feature is required or missing can be done effectively, either by practitioners or the Keras team. Tickets in keras-cv were created only when such features were NOT available or when a specific need was encountered. Commenting after a significant delay and requesting the recreation of the same issue is neither efficient nor appropriate. Kindly move this ticket (and any other similar tickets with the same context) from keras-cv to keras-hub.

[github-actions[bot]]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[github-actions[bot]]

This issue was closed because it has been inactive for 28 days. Please reopen if you'd like to work on this further.