Calibrating for Reduced Resolution

Overview

Many high-speed cameras allow speed increases by reducing (cropping) image resolution. However, calibration can be difficult or impossible at the reduced resolution; in most cases calibrating at the full sensor resolution is easier and will also give a more accurate result.

Problems with calibrating at reduced resolution

A typical high speed camera may have a resolution of 1024 x 1024 and at this resolution a standard 14x10 calibration grid, chosen to fill the field of view, will calibrate well - all coding and target dots should be recognized.

As the resolution decreases much below 1024 x 1024, the smallest dots on the grid - the two coding dots - will no longer be rendered as recognizable ellipses, instead looking more like this (greatly zoomed in):

In this case, you can still manually select the correct grid and proceed to calibrate. However, once the resolution starts to decrease more, towards 512 x 512^*, the small circles concentric with the three orientation dots will also become poorly resolved.

In this case, calibration will be impossible. To avoid this, you can calibrate at full resolution, and then perform a simple adjustment to correct for cropping.

Even if calibration is slightly possible at reduced resolution, we can often get a better result with the full field, because we can get better estimates of parameters like distortion by using data from the corners of the sensor. Because of this, it is always recommended to calibrate at full resolution, especially for critical tests.

Note: For cameras with a limited maximum resolution (IR cameras, ultra high speed cameras) we can also use special grids with sparser but larger dots - i.e., an 8 x 6 grid with very large dots. These can be generated with the target generator or you can contact Correlated Solutions to inquire about purchasing a finished grid.

Calibrating at full resolution

For full resolution calibration, set your camera for its max resolution. Speed and exposure time can be set as necessary - these will not change the calibration parameters - but aperture must not be adjusted. Black reference the cameras, if necessary; choose a grid which fills the full field of view of the sensor, and take a good calibration set.

You can then return to the reduced resolution and set up for your test - do not move the cameras or change the aperture, but lighting, FPS, and exposure time adjustment are all allowable.

Software procedure and theory

Two of the parameters we calibrate for are Center (X) and Center (Y). These are the coordinates of the pinhole center of the sensor; they tend to be roughly in the geometric center of the sensor, but never exactly, because of real-world manufacturing variation. This variation does not harm accuracy but must be calibrated for.

Vic-3D represents this as a pixel coordinate referenced to the top left of the sensor.

When we reduce the resolution - for our example, to 512 x 384 - the camera crops the image to the center of the sensor. (512, 512) is no longer the center of this reduced image - we must offset it.

To calculate this offset automatically, add both the calibration images as well as at least one speckle image (at reduced resolution) to the project in Vic-3D. Calibrate as usual, and then click File... Adjust for cropping.

Assuming the image was cropped to the center, the correct values will be filled in. Click Ok and the Center (X) and Center (Y) values will be offset as necessary. You should do this once and only once - if you click through again, the values will be offset again. Check the Calibration tab in your project - the Center (X) and Center (Y) values should be roughly in the center of your reduced resolution image.

If the image was not cropped to the optical center, you must manually enter the necessary offset values. 

This correction does not affect accuracy in any way - the digital nature of sensors means that the offset is an exact, knowable integer value.

Note: If you fail to correct for cropping, or the values are incorrect, you will most likely see a very high Projection Error in your analysis. In this case, check through the steps above and try again.

Save the project at this point, and run as usual.

 

^*All numerical values in this application note are examples and will not apply to every case.