Riddle_Training_007bbfb7 - arc-community/arc GitHub Wiki
Observations
-
I notice that the shape of the objects in the input and the output is the same, but at a different size
-
I notice that the location of the filled pixels in the output resembles the location of the pixels in the input
-
I observe that in the input, all 3x3 subblocks have a uniform colour. I classify them as "big pixels"
-
I classify this task as a "make a pattern"
-
I compare the input with the output a bit more and notice that the source of the overall pattern is the non-background pixels
-
I compare the input and the output and notice that the source of the elements for the pattern is a downsized (3x) input
Concepts Used
-
Detect background
-
Detect foreground object
-
Classify task -> "make a pattern"
-
Downscale input 3x, 2x
-
Build a pattern from a guide
-
Copy object with rescaling
Solution Approach
Possible distribution shifts
-
The background could be of a different colour
-
The source pattern could have a different colour for each "big pixel" and that could be used as a property for the resulting element
-
There could be a different operation applied to the input objects -- instead of "downsize and copy" it could be e.g. flip, rotate, hull, e.t.c.