Skip to content
All projects

Case study

Novel View Synthesis

Built an end-to-end 3D Gaussian Splatting pipeline that reconstructed a scene from 99 images, optimized 392k Gaussian primitives, and rendered RGB and depth views along interpolated camera trajectories.

Novel-view trajectory render
Contribution
Built an end-to-end reconstruction and rendering pipeline using COLMAP and 3D Gaussian Splatting.

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.