Setting Up a 3D Reconstruction Pipeline on Windows: COLMAP + Open3D + NeRF + Gaussian Splatting

This post walks through how I set up a hybrid C++/Python pipeline for 3D reconstruction and novel view synthesis on a Windows machine. The goal was to build a clean, GPU-accelerated setup for comparing Neural Radiance Fields (NeRF) and Gaussian Splatting side-by-side — starting from unordered image collections.

---

💻 System Overview

Component	Spec
OS	Windows 10 x64
GPU	NVIDIA GeForce RTX 3070
Driver	NVIDIA-SMI 561.19
CUDA	12.6 (native), 11.8 (runtime used)
Python	3.10 (via conda)
Compiler	MSVC / Visual Studio 2022

---

⚙️ Toolchain at a Glance

Tool	Use Case	Install Method
COLMAP	SfM + Dense MVS	C++ via vcpkg
Open3D	Point cloud + mesh visualization	Python via pip
PyTorch	NeRF + splatting tools	conda w/ CUDA
pycolmap	Python SfM experiments (CPU-only)	pip

---

🛠️ COLMAP Build with vcpkg (C++)

I used vcpkg to build COLMAP from source with CUDA and OpenGL enabled.

# Install vcpkg
git clone https://github.com/microsoft/vcpkg
cd vcpkg
.\bootstrap-vcpkg.bat
.\vcpkg install colmap[cuda,tests]:x64-windows

More information can be found at https://colmap.github.io/install.html

COLMAP is run via subprocesses in Python to keep the GPU-accelerated MVS backend and avoid pycolmap's limitations on Windows.

---

🐍 Python Environment (conda + pip)

conda create -n mesh-recon python=3.10 -y
conda activate mesh-recon

# PyTorch with CUDA 11.8
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia -y

# Visualization & geometry tools
pip install open3d pycolmap matplotlib numpy tqdm

# NeRF Studio
pip install nerfstudio

Note: Open3D and pycolmap install cleanly on Windows via pip. Pycolmap is CPU-only here, so COLMAP C++ remains the primary reconstruction engine.

---

🧪 Goals of This Pipeline

• Run SfM + MVS robustly on local image sets
• Preprocess outputs for both NeRF and Gaussian Splatting
• Use Open3D for quick point cloud + mesh previews
• Compare rendering quality + performance

---

🔄 Why Not Just Use Linux?

Good question. I could have used WSL2, but I wanted native GPU access and better integration with Windows-based tools — especially for testing pipelines with real-world user environments.

---

🚧 What's Next

I'll follow up with:

• Scripts to batch-run COLMAP from Python
• Conversion utilities for NeRF / splatting input formats
• Evaluation of memory/performance tradeoffs

Let me know if you’d like a copy of the environment file or build scripts.

---

Thanks for reading!