LEARNING SCENE SPECIFIC RECONSTRUCTION OF MONTE CARLO PATH TRACED IMAGE SEQUENCES WITHOUT ADDITIONAL CLEAN DATA

When using deep learning models for reconstruction of one path per pixel Monte Carlo path traced image sequences, reconstruction of unseen features can be a concern. Th\u91is can be solved by training the model on the same scene it is supposed to reconstruct images from. Learning to specialize with additional clean targets would be extremely time consuming, instead training with additional noisy targets saves time as additional noisy images is tremendously faster to render. \u91This thesis shows that a model trained without clean targets on the same scene it is reconstructing images from can under certain conditions out performa model trained on clean targets from diff\u82erent scenes. It also shows that \u80first training a model on clean targets followed by noisy targets can help with instability issues one can encounter when only training with noisy targets. Lastly, the thesis shows how there is not much of an improvement going from 400 to 1000 training examples or going from noisy targets rendered with one path per pixel to targets rendered with eight paths per pixel