When humans look at abstract art, they get a sense that each work is composed of a series of shapes, curves, lines, or general objects. We also have an intuition about how those objects would move if they pushed, pulled, or otherwise manipulated, even though we've never seen an abstract painting move before.
Our project focused on teaching computers the unspoken and intuitive shared experience of imagining interacting with abstract art. We also investigated what happened when we made art move in a precisely unintuitive way.
What it does
- First, we use KMeans and marching cubes in order to evaluate noisy counters for the image.
- Next, we smooth the counters via L1-regularized curve fitting and LASO regression.
- Next, we use region growing algorithms to section off our image into distinct color zones.
- Finally, we build a k-NN bitmap projection of the image. It looks like the original, but it is interactive!
- The resulting render is composed of curves that can be intuitively manipulated with shifts, stretches, and rotations.
In order to create anti-intuitive manipulations, we use the Fourier phase shift principle in order to change the often-assumed and rarely-modified phase space of the frequency domain.
How we built it
With machine learning, numpy, and love <3.
Challenges we ran into
Machine learning is always very difficult to get right, especially when you are rolling the algorithm by hand.
Accomplishments that we're proud of
We think the demo is pretty cool.
What we learned
We learned a ton about the current state of image analysis and standard techniques involved in forming color palettes, segmenting images, and de-noising contours.
What's next for Art-amiss
We're going to deploy it to a server so other people can try it out!
Please kindly consider us for:
- Best use of machine learning
- Best web app