Kifu: Go game record (kifu) generator

From ElphelWiki
Revision as of 03:44, 30 June 2009 by Luxigo (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Goban and Grid Detection:

1. The input image is filtered and thresholded and a big quadrilateral shape is selected, sorting contours detected with cvFindContours() from OpenCV.

Kifu region.jpg

2. The selection can be rectified with cvGetPerspectiveTransform() and cvWarpPerspective() but using a specific procedure allow to save the original coordinates in a table at the cost of a few megabytes of ram. It should be also more performant, avoiding some unused features overhead.

Kifu rectify.jpg

3. cvHoughLines2() find lines in the rectified selection and we compute the intersections for each detected line segment.

Kifu intersections.jpg

4. The intersection is first validated when the angle formed by the intersection point and the line extremities is near 90 degrees; and other detected intersections that are less distant than ~92% of the expected grid spacing are discarded. Finally each 'cell' or group is validated when the 2 opposite corners are around 90 degrees.

Kifu intersections2.jpg

5. Intersections can be mapped back to the pre-warped coordinates and vote for the original point, and we restart at step 2 with the quadrilateral found with the next threshold value.

Kifu backproj.jpg

6. Assuming extreme intersection coordinates in the original image are the corners, the transformation matrix is recomputed using those "corners" and intersections are mapped into this plane.

Kifu rectify2.jpg

7. Numbering is made line per line, starting from the top left corner, looking for the nearest neighbour on the right of the current position, up to the end of line. The next line is found looking for the nearest neighbour below the line start, etc

Kifu numbering.jpg

8. Finally, we can validate the grid simply by checking the count of numbered intersections, or restart at step 1 with another threshold and look for missing ones.

Image change detection:

Using vertical and horizontal RGB components sums of a thresholded difference between a reference image and the current one,

it is easy to compute the image coordinates of a played stone

Kifu threshold.jpg Kifu sums.jpg

Stone detection:

When a stone is played it overlaps 1 grid square on a corner, 2 on the borders and 4 in the center.

Computing horizontal and vertical pixel sums for each grid cell or around each intersection can tell where the stone is played and reveal the color of the stone, being darker or brighter than the empty intersection region.

The intersection can also be seen on the difference image when the stone is white, that could also be used to detect the stone color.

Methods for mapping the coordinates:

"2.2 Perspective transformation with two vanishing points" (pages 2 and 3, equations 7 and 10)

Inverse homography and plane image rectification (page 14)

"Inferring Projective Mappings" (page 3)

with related java and C source code here:



Photointerpretation and Small Scale Stereoplotting with Digitally Rectified Photographs with Geometrical constraints:

Vision par Ordinateur:

Projective Mappings for Image Warping: