Optic Faux

CS289 | Spring 2015 - Final Project attempting to use a Genetic Algorithm to find new and unexpected camouflage from optic flow sensors. Much like a virtual cattle guard, it is my hypothesis that static patterns can be generated to fool optic flow into thinking the exact opposite of the real motion is happening. Both the wagon wheel effect as well as Moire patterns are examples of visual exploits to the human eye, one an issue of "framerate" or scanning and the latter of "resolution." In this project, we will start by trying to fool an artificial eye, with hopes of then fooling other simple eyes. by Jonathan Bobrow, Serena Booth, Jacob Peters [source code]

[opticfaux]

Python OpenCV Optic Flow test with animated square translation. [FIXED]

square: 50 x 50px

canvas: 320 x 240px

Python Farneback settings:

flow = cv2.calcOpticalFlowFarneback(prv, nxt, 0.5, 4, 8, 2, 7, 1.5, cv2.OPTFLOW_FARNEBACK_GAUSSIAN)

Processing Farneback settings:

private double pyramidScale = 0.5; private int nLevels = 4; private int windowSize = 8; private int nIterations = 2; private int polyN = 7; private double polySigma = 1.5; private int runningFlags = Video.OPTFLOW_FARNEBACK_GAUSSIAN;

[opticfaux]

Python OpenCV Optic Flow test with animated square translation. [BUG]

square: 50 x 50px

canvas: 320 x 240px or (240,320) as opencv’s matrix would have it (rows,cols)

I was completely confused as to why the optic flow with the same settings as my Processing sketch, with the same version of OpenCV 2.4.7 would return such drastically different results. This looks more like cellular automata than it does a robust optic flow algorithm. And then it hit me. This is feedback, I am drawing the flow to the image I have instrumented to measure flow. The optic flow visualization is yielding a flow for the next frame and so on. The reason why I noticed, was because I changed the color of the flow, and the feedback became even stronger.

[opticfaux]

Processing OpenCV Optic Flow test with animated square rotation.

Testing measurement of Curl. All orange values on right side mean negative curl, in this case, confirming rotation clockwise. All green values on right side mean positive curl, in this case, confirming rotation counter clockwise.

canvas: 320 x 240px

[opticfaux]

Processing OpenCV Optic Flow test with animated square zoom.

Testing measurement of Divergence. All green values on right side mean positive divergence, which is zooming in or getting closer to an object. All orange values on right side mean negative divergence, or zooming out and getting farther from an object.

canvas: 320 x 240px

[opticfaux]

Processing OpenCV Optic Flow test with animated square.

square: 50 x 50px

canvas: 320 x 240px

[opticfaux]

Processing OpenCV Optic Flow test with animated circle.

circle: 50 x 50px

canvas: 320 x 240px