Neural compensation of residual distortion in camera calibration

Accurate camera calibration is crucial for achieving high-precision 3D measurements, particularly when using low-cost devices in long-range or complex depth environments where modeling distortion becomes significantly more challenging. Traditional calibration approaches often fail to capture focal-depth-dependent distortion due to their limited parametric structures, compromising reliability in demanding scenarios. To tackle these issues, we propose a two-step calibration process that combines geometric distortion modeling with learning-based refinement. In the first stage, we develop a Phased Distortion Attenuation Optimization (PDAO) algorithm, which gradually reduces distortion effects across varying depths. In the second stage, we introduce a neural network-based error compensation module (NN Model). It learns residual distortion patterns by taking image coordinates as input and producing undistorted coordinates based on the optimized intrinsic parameters. Experimental results demonstrate that our method significantly improves reprojection accuracy and robustness under wide-baseline and multiview settings. The proposed approach makes it particularly suitable for high-precision 3D vision applications in challenging or unconstrained environments.