Drone Photogrammetry Pipeline

From Full Arc Drone Video to Textured 3D Mesh

---

Introduction

This repository documents an open-source workflow to convert a full-arc drone video into a photogrammetry pipeline, resulting in a textured 3D mesh suitable for cultural heritage documentation and fabrication.

The workflow was developed as part of an independent research project focused on documenting twin large-scale stone labyrinths in Salem, Tamil Nadu, India, estimated to be over 1200 years old.

The emphasis of this repository is strictly technical:

• data acquisition
• reconstruction
• meshing
• export

---

Example Subject

Stone labyrinth (Labyrinth 1)
Salem, Tamil Nadu, India

---

Related Data & Outputs

Raw Scan (Unprocessed)

• Sketchfab (raw reconstruction):
  https://sketchfab.com/3d-models/ezhlu-suthu-kottai-labyrinth1-fd98e6ab4a6a447d8f8aaa80f26e390e

Processed & 3D-Printable Files

• Cleaned STL models (Rhino 8 + Bambu Studio):
  (https://sketchfab.com/3d-models/labyrinth-model-for-3dprinting-3298401c8cc6465da40a7c1e2170e6f6)

Google Drive Structure

• Full project directory (COLMAP workspace, intermediate outputs):
  https://drive.google.com/drive/folders/1OHUfLgzkVfzxqVNqoa2_PZS4i0-HLScH

Source Drone Videos

• Drone videos used for frame extraction:
  https://drive.google.com/drive/folders/1G-3tqRPTlefHTVRCwlt7l8SgC4OGmAsO

---

Publication Context

The archaeological documentation resulting from this workflow was published in:

• Caerdroia – Issue 54 (2025)
  https://labyrinthos.net/caerdroia54.html

This repository focuses only on the digital documentation and reconstruction workflow.
Interpretive, mythological, or religious narratives are intentionally excluded.

---

Overview of the Pipeline

This repository describes a fully open-source photogrammetry pipeline, implemented entirely in Google Colab, that converts drone video into a vertex-colored (textured) PLY mesh.

Core Tools Used

• FFmpeg – frame extraction from video
• COLMAP – sparse & dense photogrammetry
• Open3D – meshing, color transfer, export

Final Output

• Textured PLY mesh
• Suitable for:
  • visualization
  • archival documentation
  • downstream cleanup and fabrication

---

Requirements

This pipeline runs in a Linux / Google Colab environment.

System

• Google Colab (GPU optional but recommended)

Dependencies

sudo apt-get install colmap
sudo apt-get install ffmpeg
sudo apt-get install xvfb
pip install open3d

Dependency roles

COLMAP – Structure-from-Motion & dense reconstruction

FFmpeg – video → image extraction

Xvfb – headless execution for COLMAP in Colab

Open3D – mesh processing and file export

##Workflow

1. Frame Extraction from Drone Video

Drone video is converted into individual image frames using FFmpeg.

Copy code
ffmpeg -i input_video.mp4 -vf fps=2 -qscale:v 2 images/frame_%04d.jpg
Extracts frames at 2 FPS

Produces high-quality JPEG images

Lower FPS reduces redundancy and computation time

2. Sparse Reconstruction (Structure-from-Motion)

COLMAP computes camera poses and a sparse point cloud.

Example (automatic pipeline):

Copy code
colmap automatic_reconstructor \
  --workspace_path PROJECT \
  --image_path images \
  --quality high \
  --dense 1 \
  --use_gpu 1

In the notebook, this is executed using xvfb-run to support headless GPU SIFT in Colab.

3. Dense Reconstruction

The dense pipeline consists of:

image undistortion

patch-match stereo

stereo fusion

Key commands:

Copy code
colmap image_undistorter
colmap patch_match_stereo
colmap stereo_fusion --output_path dense/fused.ply
Output:

fused.ply

Dense point cloud with XYZ + RGB

Often contains millions of points

4. Mesh Reconstruction (Poisson Surface)

The dense point cloud is converted into a watertight mesh using Poisson Surface Reconstruction in Open3D.

Copy code
mesh, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(
    pcd, depth=11, scale=1.1
)
mesh.remove_unreferenced_vertices()
mesh.compute_vertex_normals()
Output:

mesh_poisson_o3d.ply

Smooth, continuous surface

Open3D was chosen instead of COLMAP’s Poisson mesher for greater control and extensibility.

5. Color Transfer (Textured Mesh)

Vertex colors are transferred from the dense point cloud using nearest-neighbor lookup.

Copy code
pcd_tree = o3d.geometry.KDTreeFlann(pcd)
colors = []

for v in mesh.vertices:
    _, idx, _ = pcd_tree.search_knn_vector_3d(v, 1)
    colors.append(pcd.colors[idx[0]])

mesh.vertex_colors = o3d.utility.Vector3dVector(colors)

6. Export Textured PLY

Copy code
o3d.io.write_triangle_mesh(
    "meshed-poisson-colored.ply",
    mesh,
    write_vertex_colors=True
)

Final output:

meshed-poisson-colored.ply

Vertex-colored, textured mesh

Viewable in MeshLab, Blender, or Rhino

Post-Processing (Outside Colab) While reconstruction is performed in Colab, cleanup and fabrication preparation were done using:

Rhino 8 – mesh cleanup, decimation, repairs

Bambu Studio – slicing and print preparation

Results

Fully open-source, reproducible workflow

Textured PLY mesh suitable for documentation

STL derivatives suitable for 3D printing

Applicable to large-scale outdoor heritage objects

Notes on Performance

COLMAP dense reconstruction is slow on Google Drive

Large datasets may take tens of minutes to hours

Disk I/O is a major bottleneck in Colab

Faster Alternative For production or time-critical workflows:

Agisoft Metashape

significantly faster processing

user-friendly GUI

commercial, closed-source

This repository prioritizes transparency and reproducibility over speed.

License

This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

You are free to:

share and adapt the material

use it for non-commercial purposes

Under the conditions that:

proper attribution is given

derivatives are shared under the same license

License details: https://creativecommons.org/licenses/by-nc-sa/4.0/