Abstract:
Learning scene geometry has become a potential frontier in the field of computer vision, which
uses RGB value prediction to provide important information about objects and settings. Neural
Radiance Fields NeRF is one approach in computer vision for view synthesis. It takes 5D input
parameters and outputs RGB and depth values at a certain point. NeRF applications include
robotics, mapping and Heritage. One of the key aspect of this approach is that it uses plain
neural networks to achieve this task and impact of different model depth, parameters and loss
functions on the overall performance is an area which requires further exploring. Therefore,
this thesis focuses on training and prediction of RGB values of different models starting from
shallow to deep architectures and fine tuning of hyper parameters and the impact of learning
rates on on scene generation. This thesis aims to contribute into predictive attribute learning,
paving way for enhanced video processing, video compression, video frames interpolation and
other real world applications.