The problem of this thesis is determining the scene illumination by predicting the light source for use in augmented reality. We propose a new technique for determining light maps using deep neural networks and a new synthetic dataset for model learning. Light maps represent the encoding of the two angles needed to determine the light source into a matrix of all possible pairs of them. We test and compare different neural network backbone architectures and different model learning data augmentation techniques. To test the performance, we compare the model with two previously tested methods for predicting radians and disjointed angles on a real unseen dataset. The final model achieved more accurate predictions than the previously mentioned techniques.
|
