As posted by @vlordier in the too quickly closed issue #2395 (IMHO), DCN is now a well supported feature in various frameworks. Maybe time has come to add this feature to ONNX? Is there a plan to have it in future releases?

But to generate it in ONNX it need to be converted from someplace. If this cause a massive accuracy gain, why no framework implemented it naively? Especially, this was published 2 years ago. To add an OP in ONNX, it needs to be available in one framework at least.

@ebarsoum Deformable Convolutions are also used to process equirectangular projections of omnidirectional images by parametrising the convolution kernel

More recent work Kernel Transformer Networks for Compact Spherical Convolution

Distortion-Aware Convolutional Filters for Dense
Prediction in Panoramic Images

or Corners for Layout: End-to-End Layout Recovery from 360 Images

Originally posted by @vlordier in #2395 (comment)