What I Learned About 3D Reconstruction — My First Week at PreserveMy.World

#3d #python #beginners #machinelearning

I just started my internship at PreserveMy.World, a project focused on digitally preserving cultural heritage sites using technology. This is my first week and I spent it setting up my environment and diving into 3D reconstruction research.

What is PreserveMy.World?

PreserveMy.World aims to digitally preserve real places — historical buildings, heritage sites, and landmarks — so they can be experienced even if they're damaged or lost. 3D reconstruction is at the heart of this mission.

The 5 Methods I Researched

1. COLMAP / Structure-from-Motion
Takes multiple photos of the same place from different angles and builds a 3D point cloud. Good for reconstructing buildings from tourist photos.

2. NeRF (Neural Radiance Fields)
Uses a neural network to learn how a scene looks from any angle. Produces photorealistic results but needs a strong GPU.

3. Gaussian Splatting
A newer, faster alternative to NeRF. Produces high quality real-time 3D scenes with fewer photos.

4. Monocular Depth Estimation
Estimates depth from just ONE photo. The easiest to get started with — no special equipment needed.

5. Multi-View Stereo
Uses multiple calibrated photos to build a dense 3D mesh. Good for detailed artifact models.

My Experiment — Monocular Depth Estimation on Google Colab

I decided to start with monocular depth estimation since it only needs one image and runs free on Colab.

I used the Depth-Anything-V2 model from HuggingFace to run depth estimation on a synthetic image I created with Python.

The result was really cool — the model correctly identified which parts of the image were closer and which were further away, shown through a plasma color map (yellow = close, purple = far).

What Failed

My first two attempts failed because I tried downloading images from Wikipedia URLs which got blocked. I fixed it by generating a synthetic image using Python's PIL library instead.

What I Learned

Monocular depth estimation is a great entry point into 3D reconstruction
You don't need expensive hardware to get started — free Colab works fine
For PreserveMy.World, this could be used to get quick depth previews of heritage site photos

GitHub Repo

All my code, research notes, and experiment outputs are in my repo:
https://github.com/HumnaAttique/PMW-day1

Next Steps

I want to try COLMAP next week with real photos, and eventually work toward NeRF or Gaussian Splatting for more realistic heritage site reconstructions.