PyTorch Feature Volume Rotator
This function uses trilinear interpolation to rotate and translate feature volumes. Because the function is written entirely in PyTorch, it can be seamlessly plugged into a deep learning training pipeline. As a toy example, given the following output feature volumes:
and the associated rotations that produced them, the true input feature volume (left) can be recovered from a randomly initialized feature volume (right):
python3 optimize_example.py
Examples
python3 rotation_translation_example.py rotate
produces these:
python3 rotation_translation_example.py translate
produces these: