keras-team / keras-cv

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

Porting/Refactoring TFA dense_image_warp #644

Open Luvideria opened 2 years ago

Luvideria commented 2 years ago

This ticket refers to already opened ticket #163 and https://github.com/tensorflow/addons/issues/2733

Currently tensorflow_addons.image.dense_image_warp(...) only supports bilinear interpolation, which is good enough for most, not enough for some. I implemented a higher order interpolation scheme: catmull rom (=Lanczos2) which gave better results (see: https://github.com/Luvideria/tensorflow-dense-warp-catmull-rom). It works fine but may need to be adjusted. It currently has a high maximum memory use (because of the buffer duplication for vectorization), I also managed to make another version using less memory but the code is too hacky (and slower).

I am not very familiar with building very fast or optimized tensorflow code, my implementation mostly relies on (and modifies) the previous implementation (bilinear).

The issue #163 relates to elastic transform, which can be improved by a higher order interpolation scheme.

github-actions[bot] commented 9 months ago

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.