lens-model: differentiable camera lens model

Python · PyTorch · optics · computer vision · research in progress, private repository

Classical camera calibration treats a lens as a few distortion coefficients and locates calibration markers as blob centroids. On fast or imperfect glass, that ignores the physics: aberrations and vignetting shift the apparent marker centre by pixels. This project asks how much accuracy a physically grounded model can recover.

• Ground truth from ray tracing. A real lens prescription is ray-traced with DeepLens to produce photogrammetry images with known marker positions, including a full sensor noise model.
• White-box forward model. Distortion, field-dependent Zernike wavefront, a cat-eye pupil and radiometry, all differentiable in PyTorch.
• Photometric solver. Bundle adjustment with a ZNCC curriculum followed by Levenberg-Marquardt over all parameters, evaluated on held-out poses and compared with intensity centroids, ellipse fits and OpenCV reprojection.
• Rigorous evaluation. Errors are always reported against several centre conventions (chief ray, zero-tilt, PSF centroid) so a flattering target cannot pass unnoticed. Derivations and measured findings are written up alongside the code.
• Runs are queued on GPU machines with nodes.