System information

• TensorFlow version 2.3.0 and everything latest and greatest:
• Are you willing to contribute it (Yes/No): Depends, there are already a few written by developers that might be a solution for this request.

Describe the feature and the current behavior/state.
Basically a gradient matrices before the convolutional operation (Conv1D, Conv2D, Conv3D) that gives more information about convolutional operation place in the image.

Will this change the current api? How?
I don't think so. CoordConv can be added as a separate layer before the convolutional operation.

Who will benefit with this feature?
In general, all the segmentation or detection task in which objects transition in the image is important

Any Other info.
Article: https://arxiv.org/abs/1807.03247 and original implementation by authors https://github.com/uber-research/CoordConv
Video explanation: https://www.youtube.com/watch?v=8yFQc6elePA
Possible implementation: https://github.com/mvoelk/keras_layers
Similar feature issue/mention : #32222