Why is CUDA not used for geomteric transforms?

[willcray]

The transforms in mmcv.image.geometric.py make many calls to opencv that appear to not leverage CUDA. For example, cv2.warpAffine is called here. This operation is just a matrix multiplication, which are massively accelerated by leveraging a GPU over a CPU. It not using the GPU is creating a bottleneck in my current training pipeline.

Does anyone know why CUDA isn't used or if I'm missing something? I would imagine this would massively accelerate the data augmentations in many people's training pipelines immediately.