Methodology

  • We notice that in complex indoor environments, NeRFs achieve better novel view synthesis than Gaussian Splatting (100+ FPS) but are slow to render (~4 FPS)
  • Our approach leverages a trained NeRF model to generate renders from novel viewpoints to improve scene coverage and increase the data available for training a 3D Gaussian Splatting model.
  • We train a Nerfacto instance to generate novel viewpoints and utilize Scaffold-GS as our 3D Gaussian Splatting framework.

Figure: A high level flow-chart depicting our methodology


  • To densify the number of views available for training, we sample synthetic views from the trained NeRF model in a hemisphere centered towards the table.
  • We aim to explore efficient sampling strategies for synthetic images that enhance reconstruction quality while minimizing the number of additional synthetic images required.

Figure: A set of cameras showing the sampled novel viewpoints to augment the training dataset