3D Scene Reconstruction using Structure-from-Motion and NeRF CSCI

Ignacio Blancas Rodriguez, Ryan Klivansky, Luke Riley, Andrew Devon Boyacigiller 1430
Motivation Problem Goal
3D Scene Reconstruction from uncalibrated Use SfM and NeRF to reconstruct 3D scenes
Traditional methods for 3D Reconstruction methods like
from uncalibrated image sequences
images scene reconstruction from photogrammetry also
1. Estimate camera poses w/ Colmap, cast rays through each
image sequences often require ideal diffuse lighting
• High fidelity novel view generation struggle with accurately and a large number of high-res
pixel
2. Render the volume as a Radiance Field for reconstruction
• Low amount of image input capturing fine geometric images. 3. Generate novel views of the scene from new camera poses
needed/accessible for non-photographers details and
photorealistic lighting effects.

What is the insight into the problem that leads to the solution? Pipeline Overview

Results (images/figures)
Model Architecture
Beginning Density Estimation Color Prediction

γ(x)
63
256 256 256 256 256 256 256 128 3

γ(x)
*While the novel views are present in these
images, they ended up tiled in a 5x5 pattern.
This did not appear in the above scenes, whose γ(d)
images are made for NeRF reconstruction. See
our report on image dimensions and model
architecture.
63 27
References Acknowledgements
[1] Mildenhall, Ben, et al. "Nerf: Representing scenes as neural radiance fields for view synthesis." Communications of the ACM 65.1 (2021): 99-106.
[2] "COLMAP – Structure-from-Motion and Multiview-Stereo" https://demuc.de/colmap/
We would like to thank Prof,. Sridhar, as well as all the TAs for allowing us to have
[3] NVLabs. "colmap2nerf" https://github.com/NVlabs/instant-ngp/blob/master/scripts/colmap2nerf.py such a wonderful semester. We'd also like to thank our mentor, Nwo Wen Lei.