Rubiksolver.Demo.small.mp4
Rubiksolver is a simple GUI-based computer vision application that detects and solves a Rubik’s Cube in real time.
I use Canny edge detection to extract individual facelets from the video frame. To reliably distinguish cube facelets from background objects, the program only accepts contours that meet specific aspect-ratio and size thresholds, which are adjustable from the GUI.
To further eliminate erroneous facelet candidates, I project the candidates onto a grid and select those with the lowest error after 100 RANSAC iterations.
Finally, to correctly label each facelet by its face color, I use the aggregate mean hue and saturation of its pixels to classify colors. I found this model to be fairly robust and invariant to lighting changes compared to RGB-based approaches.
The figure above shows the aggregate mean hue and saturation of each facelet’s pixels across multiple runs under different lighting conditions. As you can see, a very simple thresholding model suffices to classify each facelet by color.
To run the app, you’ll need to have uv
installed on your system.
If you don’t already have it, follow the
installation instructions here.
git clone https://github.com/pdadhikary/rubiksolver.git
cd rubiksolver
uv syncThe above commands will clone the repository and install all required dependencies in a virtual Python environment. Once setup is complete, you can start the program with:
uv run rubiksolverWhen scanning their Rubik's cube the user should hold up each face of the cube to the webcam. By convention we assume that the white face is UP and the yellow face is DOWN. When scanning the white face, the red face is DOWN and it should be UP when scanning the yellow face.
Once the scan is complete press the Play button to animate the solution steps.
You can also step through each of the moves using the Previous and Next buttons.
- Denoise Filter Diameter Size of the denoise filter kernel.
- Color Filter Sigma The colorSigma parameter for the
bilateral filter.
A larger value means more smoothing
of the image as pixels with more
different color values are averaged
together. A smaller value correlates
to less smoothing and edges are better
preserved.
- Space Filter Sigma The spaceSigma parameter of the
bilateral filter.
A larger value means more smoothing as
a wider neighbourhood of pixels is
considered. Whereas a smaller value
give a more localized smoothing.
- Canny Threshold The upper and lower thershold for the
Canny edge detection hysteresis step.
Lower values correspond to the weaker
edges present in the image and higher
values represent more pronounced edges.
- Facelet Area Threshold Size thershold of candidate facelets
contours.
This option can be used to eliminate
erroneous contours from the background
by eliminating contours that are either
too small or too large. The size is the
percentage of screen space.
- Facelet Contour Area Ratio Threshold The width to height ratio of contours
to consider as candidates.
This value can be changed to elimate
contours that are not square-ish.
- Facelet Bounding Box Aspect Ratio Threshold Metric for how well the contour
fits/occupies the bounding box.
This can be changed to allow for
facelets that are not perfectly square,
i.e, may have rounded edges or are
circular.
- RANSAC Error Threshold Metrix for how well the detected
contours fit in a 3x3 grid.
Used to filter out outlier/erroneous
contours that may be present in the
background.