Distorted image sections

[pi314ever]

Currently, image distortion appears very choppy (especially apparent when compared against the naive approach).

Potential improvements:

• Retrain neural model multiple times over a larger range
• Increase model size (Maybe max limit to fit in GPU cache)
• Use an algorithm that optimizes hyperparameters on undistorted --> distorted --> undistorted image difference
• Use an iterative algorithm as baseline and optimize based on that

[pi314ever]

Other potential improvements:

• Split up model into radial and tangential components to make it easier to train
• Somehow encourage radial symmetry for radial distortions (by using polar coordinates?)