Self-Calibration Toolbox

NINOX 360 has developed the ability to calibrate camera systems composed of one or more cameras using natural features alone. No calibration grids or known structures are required. It enables recalibration on the fly in the field without infrastructure. Need to replace a lens but don't have a replacement with the same field-of-view? No problem, camera characteristics are determined without priors and it informs you when it has converged to stable solution.

This works using either motion in the scene or the camera and by applying geometric constraints, combined with machine learning. No retraining is required for different environments, camera technologies, and minimal priors are required. Hard mathematical evidence is used to determine the quality of provided estimates.

NINOX 360 is looking for beta testers in difficult industries for this toolbox. Please fill in the form below or contact sales@ninox360.com for more information.

Key Words: Camera Calibration, Self Calibration, Auto Calibration, in situ calibration

Inputs

  • Multiple images of the same scene

  • Video Feed

  • Pre-computed associated image features

Outputs

  • Best estimate of calibration parameters and camera model

  • Estimated variance of parameters

  • Quality of data report

  • Suggestion for where to collect new views

  • Elapsed time: Typically 10 sec to 5 min depending on amount of data

Real World Results

Chessboard

Mean Residual Error: 0.291 pixels

Focal Length: 568 px

Radial Distortion: [ 0.03116 , -0.045 ]

Self Calibration with Natural Features

Mean Residual Error: 0.856 pixels

Focal Length: 630 px

Radial Distortion: [ 0.1643 , -0.1217 ]

Using a known target tends to result in more accurate calibration immediately after calibration. However, degradation over time, temperature, and mechanical shock will lend itself towards a system capable of recalibration.

Technical Explanation

In 2019 a team member describes an early version of this technology where it successfully calibrates and creates a 3D reconstruction from just three images. While an impressive demonstration, it lacks capability to determine if it converged to a valid solution and uses what is now out dated approaches.