How 3D Reconstruction Is Helping Save the World's Cultural Heritage — My Week 1 at PreserveMyWorld

Copy everything below:

I come from a background of building things — ML pipelines on 3 million PUBG records, swarm robotics simulations, IoT robots, and 2D games in C++. But this week, starting my internship at PreserveMyWorld, I encountered something that connected all of that to a much bigger mission: using AI to make sure humanity's physical heritage is never truly lost.
This is my Week 1 write-up from the PMW AI Track — what I learned about 3D reconstruction, why it matters, and what I'm planning to experiment with next.
What Is PreserveMyWorld?
PreserveMyWorld (PMW) is a project focused on digitally documenting and preserving cultural heritage sites — monuments, ruins, and landmarks that are threatened by conflict, climate change, or neglect. The AI track I joined is focused on using computer vision and 3D reconstruction to create high-fidelity digital twins of these sites.
The Three Methods I Explored

1. Structure from Motion (SfM) — The Foundation How it works: You feed overlapping photographs into a pipeline like COLMAP. The algorithm finds matching keypoints across images, then simultaneously solves for the 3D positions of those points and the positions of the cameras that took the photos. The output: a 3D point cloud of the scene and a precise map of where every camera was. Why it's powerful for heritage: You don't need special equipment. Researchers have used ordinary tourist photos already uploaded to the internet to reconstruct monuments that were later destroyed. The data was always there — SfM extracts the 3D structure from it. What I want to try: Running COLMAP on a set of photos I take of a building near my campus in Islamabad.
2. Neural Radiance Fields (NeRF) — The Neural Revolution How it works: NeRF trains a small neural network to represent a scene as a continuous 3D volume. Given any camera position and viewing direction, the network predicts the color and density at every point along that ray. You render images by casting rays through this volume — a process called ray marching. The result: Photorealistic novel-view synthesis — you can render the scene from angles that were never in the training photos, with accurate lighting, soft shadows, and fine surface detail. Tools: Nerfstudio is the go-to open-source toolkit. It packages multiple NeRF variants into a clean, modular Python library with great documentation. Why it matters for PMW: A NeRF of a heritage site is not just a 3D model — it is a photorealistic, fully explorable memory of that place at a specific moment in time.
3. Gaussian Splatting — The Real-Time Frontier How it works: Instead of a neural network, Gaussian Splatting represents a scene as millions of tiny 3D Gaussians — small ellipsoidal blobs of color. These are optimized directly from photos and rendered in real time by projecting them onto the image plane. The big win: Real-time, interactive rendering. NeRF can take seconds per frame; Gaussian Splatting runs at 30–100+ FPS. You can walk through a scene interactively in the browser. Why it matters for PMW: This is what makes truly accessible digital heritage possible. Not a video tour — an interactive world you can explore yourself. Why This Connects to My Background Coming from data science and robotics, I see immediate connections:

SfM is fundamentally an optimization problem — the kind of numerical reasoning I've practiced with ML pipelines.
NeRF is a neural network trained on a custom dataset — the same workflow I know from Scikit-learn and Pandas, just with a very different target.
Gaussian Splatting involves 3D geometry and real-time rendering — the spatial reasoning I built while working on swarm robotics and sensor systems.

The tools are new, but the thinking is familiar. That gives me confidence.
What I'm Planning Next

Run a NeRF on a public dataset using Nerfstudio and document the training outputs
Try COLMAP on my own photos to generate a point cloud
Explore a Gaussian Splatting Colab notebook and compare output quality to NeRF
Write up a method comparison as a structured experiment note in my PMW-day1 repo

Closing Thoughts
Week 1 has made one thing very clear: 3D reconstruction is urgent, not just interesting. The methods are maturing fast, the open-source tools are excellent, and the mission at PreserveMyWorld gives this work real stakes.
I am genuinely excited to go deeper. If you're interested in AI for cultural preservation — or just in NeRF and Gaussian Splatting — feel free to connect.

Taha Abid — Data Science @ FAST NUCES, Islamabad

PreserveMyWorld AI Track Intern · Week 1 · June 2026

📧 [e-mail] · LinkedIn: linkedin.com/in/taha-abid-aa9023383 · GitHub: github.com/Taha6497-eng