Serverless RunPod endpoint that runs COLMAP structure-from-motion on drone or handheld MP4 video. Part of the RunSplat pipeline.

Input: one or more MP4 video URLs
Output: COLMAP workspace (base64 tar.gz) ready for 3D Gaussian Splatting training

MP4 video(s)
    │
    ▼  ffmpeg — extract up to 150 frames
    │
    ▼  COLMAP feature_extractor
    │
    ▼  COLMAP sequential_matcher / exhaustive_matcher
    │
    ▼  COLMAP mapper (sparse reconstruction)
    │
    ▼  COLMAP image_undistorter
    │
    ▼  colmap_workspace.tar.gz (base64)

Traditional photogrammetry outputs meshes or point clouds. Gaussian Splatting is different: it lets you fly through the scene and see it exactly as it is — from every angle — with real textures.

• See your land as it really looks, not an abstraction
• Easier to understand than technical maps or wireframes
• Ideal for design reviews, planning, and client presentations
• Take remote measurements without revisiting the site
• No heavy software needed — runs in any modern browser
• Track construction progress or erosion over time

COLMAP is the industry-standard open-source Structure-from-Motion tool that computes camera poses from overlapping images, producing the sparse point cloud and calibrated cameras that Gaussian Splatting trainers like Brush need as input.

For best COLMAP reconstruction results, circular flight patterns significantly outperform traditional grid surveys:

Unlike traditional grid or double-grid patterns, Circlegrammetry enables drones to fly in circular patterns, with the camera angled between 45° and 70° toward the center of each circle. This method captures images from more angles in fewer flights.

— SPH Engineering, Circlegrammetry

More viewing angles = more robust feature matching = better COLMAP reconstruction = better Gaussian Splat.

{
  "input": {
    "video_url": "https://example.com/drone-flight.mp4",
    "num_frames": 150,
    "matching": "sequential",
    "camera_model": "SIMPLE_RADIAL",
    "sequential_overlap": 10,
    "quality": "high",
    "gpu": true
  }
}

Multiple videos (combined into one reconstruction):

{
  "input": {
    "video_urls": [
      "https://example.com/pass-north.mp4",
      "https://example.com/pass-south.mp4"
    ],
    "num_frames": 200,
    "matching": "sequential"
  }
}

{
  "colmap_workspace_b64": "<base64 tar.gz of colmap/ directory>",
  "colmap_version": "4.0.3",
  "num_frames_extracted": 147,
  "status": "done"
}

The colmap_workspace_b64 field contains a gzipped tar archive of the colmap/ directory, which includes:

colmap/
├── database.db          # Feature match database
├── images/              # Undistorted frames (Brush input)
└── sparse/0/            # Camera poses + sparse point cloud
    ├── cameras.bin
    ├── images.bin
    └── points3D.bin

Pass this directly to brush-serverless as colmap_workspace_b64.

• Docker with NVIDIA Container Toolkit
• NVIDIA GPU with CUDA 12.9 driver support

# First build compiles COLMAP 4.0.3 from source — allow ~20–30 min
docker build -t colmap-serverless:local .

# Stream output and save log
docker build -t colmap-serverless:local . 2>&1 | tee /tmp/build.log

# Filter for errors only
docker build -t colmap-serverless:local . 2>&1 | grep -E "^(ERROR|Step|error|failed)"

./scripts/test_local.sh

# Skip rebuild if already built
./scripts/test_local.sh --no-build

The test downloads lighthouse.mp4, extracts 30 frames, runs the full COLMAP pipeline, and validates that colmap_workspace_b64 is present and non-empty in the output.

docker run --rm --gpus all colmap-serverless:local \
  python3 handler.py --test_input '{
    "input": {
      "video_url": "https://github.com/alx/runsplat/releases/download/v0.1.5/lighthouse.mp4",
      "num_frames": 30,
      "matching": "sequential",
      "gpu": true
    }
  }'

Test that COLMAP compiles without running the handler:

docker build --target colmap-builder -t colmap-builder . 2>&1 | tee /tmp/build-colmap.log

1. Verify the local build and test pass
2. Push to GitHub and create a release: gh release create v1.0.0 --generate-notes
3. RunPod Hub detects the release, builds the image, and runs .runpod/tests.json
4. After tests pass, submit for review on the Hub page

Configuration is in .runpod/hub.json. Available presets:

• Standard drone video — 150 frames, sequential matching, SIMPLE_RADIAL
• Wide-angle / fisheye drone — 150 frames, sequential matching, OPENCV_FISHEYE
• Dense / thorough reconstruction — 300 frames, exhaustive matching

• COLMAP documentation
• Feature extraction options
• Matching strategies
• Camera models reference
• GitHub repository

• COLMAP — Structure-from-Motion and MVS
• RunPod — Serverless GPU infrastructure
• SPH Engineering — Circlegrammetry technique