Computer vision
Buildings Built in Minutes: structure from motion and NeRF
Classical reconstruction of WPI Unity Hall and neural rendering
- Timeline
- Feb 2025 to Mar 2025
- Affiliation
- WPI RBE/CS 549 Computer Vision
- Team
- Divam Trivedi
- Reprojection error
- 1.75 to 1.01 px
- PSNR vs baseline
- +60%
- Rendering artifacts
- -40%
What it is
Turns uncalibrated photographs into a metrically reliable 3D asset with a full structure-from-motion pipeline, then implements NeRF end to end to render photorealistic 360-degree orbits of the Lego, Hotdog and Ship scenes.
Problem
Reconstructing a building from phone photos for AR and VR needs both classical geometry and neural rendering while keeping compute modest.
Approach
Phase 1 in Python and OpenCV: SIFT keypoints and matching, eight-point fundamental and essential matrix estimation with RANSAC, PnP camera poses, linear and nonlinear triangulation and bundle adjustment. Phase 2: NeRF with positional encoding and hierarchical coarse-to-fine volume sampling, evaluated with PSNR and SSIM.
Highlights
Average reprojection error improved from 1.75 to 1.01 pixels across five views after sparse bundle adjustment in SciPy.
NeRF novel views reached about 60% higher PSNR than the baseline with roughly 40% fewer rendering artifacts; coordinates aligned to SfM poses for consistent scale.
Reproducibility guards: fixed seeds, deterministic loaders, metric scripts and result caching.
