Context
The problem
Reconstruct a scene from images and render consistent novel RGB and depth views along a virtual camera path.
System
Architecture
- 99 source images
- COLMAP camera-pose reconstruction
- 3D Gaussian Splatting scene representation
- 392k optimized Gaussian primitives
- Interpolated camera trajectories
- Aligned RGB and depth renders
Engineering judgment
Technical decisions
Recover calibrated camera poses before training
Extracted video frames and reconstructed camera poses with COLMAP to establish calibrated views of the scene.
Represent the scene with 3D Gaussian Splatting
Trained a 3D Gaussian Splatting representation that optimized 392k Gaussian primitives from the reconstructed image set.
Render aligned RGB and depth trajectories
Rendered RGB and depth views along interpolated camera paths so the trajectory output shows both appearance and scene depth.
Result
Outcome
Produced side-by-side novel-view trajectory renders with RGB on the left and depth visualization on the right.