Original build
A Jupyter notebook using NumPy, OpenCV, and Matplotlib. The work includes calibration, rectification, reprojection, and failure analysis.
Project 04
Fisheye Calibration
A camera calibration workflow that uses checkerboard observations to estimate fisheye lens parameters, rectify distorted frames, and evaluate the result with reprojection error.
- Completed
- 2024
- Dataset
- 26 checkerboard frames
- Library
- OpenCV fisheye
- Valid mean error
- 12.66 pixels
Calibration walkthrough
Inspect the frame before and after rectification.
These are the actual outputs saved in the original notebook.
Collect varied checkerboard views
The lens model needs the checkerboard at different positions and angles. Frames where the pattern is incomplete or too distorted weaken the calibration.
Detect and refine corners
findChessboardCorners locates the pattern, then cornerSubPix improves the coordinates before calibration.
Drag to compare the original fisheye projection with the OpenCV rectification.
12.66pixel mean error across valid samples
Check reprojection error
OpenCV projects the known checkerboard points back into each image. The distance from the detected corners measures calibration quality. Frames 3 and 17 were excluded from the valid-sample mean after producing weak results.
Method
The notebook follows the full calibration loop.
- Object pointsDefine the checkerboard in world coordinates.
- Image pointsDetect the corresponding pixel locations.
- Estimate K and DSolve intrinsics and distortion coefficients.
- Remap pixelsGenerate and apply the rectification maps.
Why calibration matters
Wide-angle cameras provide useful field of view, but geometry-dependent tasks need a known camera model before measurements or localization are trustworthy.
Browser presentation
The page does not recompute a 26-frame calibration. It exposes the exact notebook artifacts in a focused step-by-step comparison.
Open the notebook ↗