• Introduction
• Usage
• Download
• Calibration Grids
• Questions and Feedback
• References

Detecting the sub-pixel coordinates of image features, such as the corners of a calibration grid, is the first stage any camera calibration algorithm. This grid detector was built as part of the metric calibration pipeline of the Stanford Multi-camera array which consists of 100 cameras. However, this has been used in several other projects in our lab and by some of our colleagues elsewhere, such as:

• Planar alignment via a homography for high speed videography
• Projector-camera alignment in air and underwater for synthetic aperture confocal imaging
• Automatic detection of a Macbeth color chart for color calibration (tech report)

The key strengths of our grid detector are:

• Automatic: The detection of grid features is fully automatic. No clicking of corner points by the user (as in the popular camera calibration toolbox [1]) or parameter tuning is required.
• Robust: The grid detector is robust to partial occlusion, and works even when 30% of the grid is occluded (unlike OpenCV [2]). This is useful for calibrating large camera networks since the calibration grid does not have to be completely visible in all cameras.
• Geometric verification: The grid features detected in the image can be matched against the known geometry of the actual calibration grid used. This eliminates all mismatches.

We have successfully used the grid detector on images from different kinds of cameras (including those from our array, a Canon EOS D10, an Olympus 3000Z, Sony DKC-ST5, consumer webcams) on over 10,000 images.

findgrid:
Linux binary executable, download here.

image file:
File containing image you want to run the grid detector on. PPM, PGM, PNG, JPG formats are supported.

grid file:
File name in which the output of the grid detector is written. The output grid file contains the (x,y) coordinates of the corners of grid squares detected, one corner per line. The line format is:

corner-number x-pixel coordinate y-pixel coordinate

Here is an example of a grid file. The corner numbers start at 0, and increase in row-major order. For the 6x6 grid in the example above, there are 6 x 6 = 36 squares and hence 36 x 4 = 144 corners. The top left corner is numbered 0 and the bottom right is 143.

config file:
File containing configuration parameters for the grid detector, download here. To be edited only by developers.

-e|c (optional):
This flag controls the edge detector used. -e results in the use of the Palmer edge detector, which is the default. -c results in use of the Canny edge detector. This is recommended only if the grid squares are large (more than 100 pixels x 100 pixels) in the image or axis aligned. I suggest you try both and go with whichever works better for your images.

-p side space rows cols (optional):
Specifying the geometry of the calibration grid used for verification. side refers to the length of the side of each of the grid squares, space is the distance between two adjacent squares in the same row or column. It is necessary to use the same unit of length for both. rows and cols are the number of rows and columns in the grid.

threshold: (optional)
When the calibration grid geometry is specified with -p, a 2D projective mapping (also called a homography or collineation) is computed between the detected grid squares and the specified calibration grid geometry and the RMS error (in pixels) of the mapping is computed. The user supplied threshold is used to verify the correctness of the grid detection. If the grid has been correctly detected, the RMS error should be small and below the threshold.

In rare cases, the grid detector can miss an entire row or column of squares. This is where the geometry verification is useful: the RMS error is high, and should exceed the threshold, in which case the program will print an error message and no grid file will be written. The default threshold is 1.0 pixel, for correctly detected grids the RMS error averages around 0.35 pixels.

Guidelines

• The image of the calibration grid should occupy an area at least 320 x 320 pixels. If the image of the grid is smaller, the accuracy of feature detection will diminish and the detector will begin to fail.
• We recommend downsampling high-res images to about 1 megapixel or less. For very high-res images, the number of features (edges, lines, junctions) detected may exceed allocated memory causing a segmentation fault.
• It is strongly recommended to use the geometric verification with -p unless there is a good reason not to, such as severe radial distortion.

• Please read the license and terms of use before downloading grid detector.
• If you find this program useful in your research, please cite this page and author in publications it was used in. Thank you!

There are two files you need to download:

1. the linux binary findgrid
2. the parameter config file params.cfg

[2] Phil McLauchlan. Horatio: a computer vision and image processing library.

[3] Phil McLauchlan. Gandalf: a computer vision and numerical algorithm library.

[4] OpenCV computer vision library.

© Vaibhav Vaish
Last update: January 5, 2006 11:38:23 PM