Experimental Validation of Bias in Checkerboard Corner Detection

The sub-pixel corner refinement algorithm in OpenCV is widely used to refine checkerboard corner location estimates to sub-pixel precision. This paper shows using both simulations and a large dataset of real images that the algorithm produces estimates with significant bias and noise which depend on the sub-pixel corner location. In the real images, the noise ranged from around 0.013 px at the pixel centre to 0.0072 px at the edges, a difference of around $1.8\times$. The bias could not be determined from the real images due to residual lens distortion; in the simulated images it had a maximum magnitude of 0.043 px.