Improving the convergence speed of NeRFs with depth supervision and weight initialization

Neural rendering is a new and developing field where computer graphics and deep learning techniques are combined to generate photo-realistic images using deep neural networks. In particular, Neural Radiance Fields (NeRF) is able to synthesise novel views of a scene with unprecedented quality by fitting a Multi-Layer Perceptron (MLP) to RGB images. However, training this network requires plenty of time and computation even on modern GPUs, making this new technology hardly employable on practical specialized applications. In this project, we show that employing the known depth of the scene as an additional supervision during the training, and starting from pre-trained weights of other scene with similar setups, instead of from scratch, leads to a convergence speed 3 to 5 time faster.