About This Page
This page collects my strongest work from the PMW internship: a Structure-from-Motion debugging case study, a deployed interactive AR demo, and the process evidence behind both. It's meant to serve as the final evidence pack for certificate review and as a shareable summary of what I actually built.
Track & Focus
Track: AI-Based 3D Scene Reconstruction / Platform & Web Engineering
Team: Charbagh Collective
Internship: PreserveMy.World × TechRealm — Cultural Heritage Preservation
I worked on 3D reconstruction of Pakistani heritage sites using COLMAP (Structure-from-Motion) and Three.js (interactive AR/visualization), applied to Shalimar Gardens and Lahore Fort, Lahore.
Strongest Artifacts
1. Case Study: Fixing a Failed COLMAP Reconstruction (Shalimar Gardens)
Diagnosed and fixed a failed Structure-from-Motion reconstruction of the Moorcroft Pavilion — improved camera registration from 2/15 to 15/15 and sparse point count from 42 to 842 by curating tightly-overlapping image sets.
- Write-up: DEV.to Case Study
- Repository: shalimar-gardens-3d-reconstruction
- Process evidence: Commit history
2. Lahore Fort — Interactive Three.js AR Demo
Built and deployed an interactive AR viewer using MiDaS-generated depth maps, validated against the COLMAP South Building benchmark dataset.
- Live demo: Lahore Fort AR Demo
- Repository: Lahore-Fort-3D-Reconstruction
- Process evidence: Commit history
3. Additional Technical Writing
- From 42 Points to 842: Debugging a CPU-Only 3D Reconstruction Pipeline
- Shalimar Gardens, Lahore: A Heritage Overview
What I Learned
The biggest technical lesson from this internship: in Structure-from-Motion pipelines, image overlap density matters more than image quantity. A failed reconstruction isn't always a hardware or software problem — sometimes it's a data curation problem, and the fix is going back to source images with a more deliberate eye for consistent overlap between viewpoints.
Beyond the technical work, I learned to document a debugging process transparently — showing the failed attempt (V1) alongside the fix (V2), rather than only presenting the polished final result. That transparency is what makes evidence credible for review.
Next Steps
- Apply the same image-overlap curation approach proactively to future reconstructions, rather than after a failed first attempt
- Extend the Lahore Fort AR demo with additional depth-mapped viewpoints
- Explore mesh generation from the densified Shalimar Gardens point cloud
PMW × TechRealm Internship — Charbagh Collective — August 2026
Top comments (0)