Calibration and Lens Distortion
Straight building edges sometimes curve near the edge of a photograph.
Calibration and Lens Distortion explained visually
Straight building edges sometimes curve near the edge of a photograph. English captions are optional and start off. The complete transcript follows the written lesson.
What you will learn
Straight building edges sometimes curve near the edge of a photograph.
Separate camera properties from pose
Straight building edges sometimes curve near the edge of a photograph. Lens distortion can cause that effect, and it matters when image coordinates support measurements. Camera calibration estimates parameters that connect the camera model to observed pixels. Intrinsic parameters describe properties such as focal lengths and principal point; extrinsic parameters describe the camera's pose relative to a chosen world frame.
Gather varied observations
A common calibration procedure observes a target with known geometry from several orientations and positions. Variety is important. Repeating nearly the same view provides less useful geometric information than covering different parts of the image with different tilts. The target dimensions must be measured correctly, and detected target points must correspond to the intended physical locations.
Check calibration quality
Reprojection error compares observed image points with points predicted by the estimated model. A low error is useful evidence, but it is not the only check. Inspect coverage across the image, stability on held-out views, and whether the model fits the camera. Recalibration may be needed if focus, zoom, resolution, or mounting conditions change. Save the calibration together with its camera settings and date so that it is not accidentally applied to unrelated images.
A worked example
Suppose a known target point is observed at pixel one hundred, eighty, while the calibrated model predicts one hundred and three, eighty-four. Its Euclidean reprojection error is five pixels, because the horizontal and vertical errors are three and four pixels. Report aggregate errors as well as problematic individual views.
Worked example
Suppose a known target point is observed at pixel 100, 80, while the calibrated model predicts 103, 84. Its Euclidean reprojection error is five pixels, because the horizontal and vertical errors are three and four pixels. Report aggregate errors as well as problematic individual views.
Quick check: test your understanding
Try it yourself: Why might a calibration made at one zoom setting be unsuitable after changing the zoom?
Check the reasoning
Zoom can change the effective focal length and other imaging properties, so the old intrinsic parameters may no longer describe the camera.
Does a low fitting error prove perfect real-world measurements?
No. Target errors, limited view coverage, model mismatch, and measurement conditions can still matter.
How would you solve the practice case?
Zoom can change the effective focal length and other imaging properties, so the old intrinsic parameters may no longer describe the camera.
What does the worked example show?
Suppose a known target point is observed at pixel 100, 80, while the calibrated model predicts 103, 84. Its Euclidean reprojection error is five pixels, because the horizontal and vertical errors are three and four pixels. Report aggregate errors as well as problematic individual views.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 3 transcript239 words · Video transcript · Matches the published narration and captionsOpen transcript
Straight building edges sometimes curve near the edge of a photograph. Lens distortion can cause that effect, and it matters when image coordinates support measurements. Camera calibration estimates parameters that connect the camera model to observed pixels. Intrinsic parameters describe properties such as focal lengths and principal point; extrinsic parameters describe the camera's pose relative to a chosen world frame.
A common calibration procedure observes a target with known geometry from several orientations and positions. Variety is important. Repeating nearly the same view provides less useful geometric information than covering different parts of the image with different tilts. The target dimensions must be measured correctly, and detected target points must correspond to the intended physical locations.
Reprojection error compares observed image points with points predicted by the estimated model. A low error is useful evidence, but it is not the only check. Inspect coverage across the image, stability on held-out views, and whether the model fits the camera. Recalibration may be needed if focus, zoom, resolution, or mounting conditions change. Save the calibration together with its camera settings and date so that it is not accidentally applied to unrelated images.
Suppose a known target point is observed at pixel one hundred, eighty, while the calibrated model predicts one hundred and three, eighty-four. Its Euclidean reprojection error is five pixels, because the horizontal and vertical errors are three and four pixels. Report aggregate errors as well as problematic individual views.