GeoCalib on Replicate

Single-Image Camera Calibration with Geometric Optimization

Paper: ECCV 2024 | Authors: Veicht, Sarlin, Lindenberger, Pollefeys (ETH Zurich / Microsoft Research) | Code: github.com/cvg/GeoCalib

GeoCalib estimates camera intrinsics (focal length, field of view, lens distortion) and gravity direction (roll and pitch) from a single image — no calibration pattern, no multi-view setup, no EXIF data needed.

It works on both outdoor scenes (streets, buildings, landscapes, aerial views) and indoor environments (rooms, hallways, offices, warehouses).

Default Example

Default input image:

https://cdn.sanity.io/images/k55su7ch/production2/d9e35a73891d43ccb0bc665bf2e0d5d9d6f1ea2b-4200x2363.jpg?w=1920&q=75&auto=format

import replicate

output = replicate.run("visionaix/geocalib", input={
    "image": "https://cdn.sanity.io/images/k55su7ch/production2/d9e35a73891d43ccb0bc665bf2e0d5d9d6f1ea2b-4200x2363.jpg?w=1920&q=75&auto=format",
    "camera_model": "pinhole",
    "return_visualization": True,
})

Default output:

{
  "camera": {
    "image_width": 1920,
    "image_height": 1080,
    "focal_length_x": 897.49,
    "focal_length_y": 897.49,
    "principal_point_x": 960.0,
    "principal_point_y": 540.0,
    "vertical_fov_deg": 62.07,
    "horizontal_fov_deg": 93.86,
    "intrinsic_matrix_K": [
      [897.49, 0.0, 960.0],
      [0.0, 897.49, 540.0],
      [0.0, 0.0, 1.0]
    ]
  },
  "gravity": {
    "roll_deg": 0.322,
    "pitch_deg": -0.528,
    "gravity_vector_3d": [-0.00561, -0.99994, -0.00922]
  },
  "uncertainty": {
    "roll_uncertainty_deg": 0.630,
    "pitch_uncertainty_deg": 0.936,
    "focal_uncertainty_px": 222.37,
    "vfov_uncertainty_deg": 17.74,
    "gravity_uncertainty": 0.0168
  },
  "performance": {
    "inference_time_seconds": 1.39,
    "device": "cuda"
  }
}

Indoor vs. Outdoor

GeoCalib is trained on diverse datasets covering both indoor and outdoor environments:

Outdoor: Urban streetscapes, architectural photography, landscapes, aerial/drone imagery

Indoor: Office spaces, hallways and corridors, warehouses/factories, residential rooms

The model extracts gravity cues from any vertical or horizontal structures. Confidence maps automatically downweight textureless regions (sky, uniform walls) and focus on structurally informative areas.

Inputs

Camera Models

Output

• calibration_json — Full calibration data (intrinsics, gravity, uncertainty, performance)
• visualization — 6-panel diagnostic PNG (input + fields + confidence + summary)

Advanced Usage

Known Focal Length Prior

output = replicate.run("visionaix/geocalib", input={
    "image": "photo.jpg",
    "focal_length_prior": 1500.0,
})

Fisheye Lenses

output = replicate.run("visionaix/geocalib", input={
    "image": "fisheye.jpg",
    "camera_model": "simple_divisional",
    "use_distorted_weights": True,
})

Citation

@inproceedings{veicht2024geocalib,
  title     = {GeoCalib: Single-Image Calibration with Geometric Optimization},
  author    = {Veicht, Alexander and Sarlin, Paul-Edouard and Lindenberger, Philipp and Pollefeys, Marc},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2024},
}

License

This endpoint wraps GeoCalib. See the original license.