I uploaded a random sampling of 108 “bad” compositions to Twitter, following the “good” compositions from this post using the same A-HOG data set. The “bad” set has a marginally lower mean number of likes (0.52), but more than double the mean retweets (0.44). The total number of likes for the “bad” set was 56 (compared to 68 for the “good” set); the total number of retweets for the “bad” set was 48 (compared to only 19 for the “good” set). Of course an uncontrolled variable is the size of the growing twitter audience for the Zombie Formalist. Following is a plot analogous to this post. I’ve also included the compositions from this set with the most likes and retweets (corresponding to the 5 peaks below)
While I did notice that some of the “bad” compositions were actually quite good (false negatives, as discussed in this post) this set above strikes me as quite weak. This confirms that my own aesthetic classes are not universal (this is obvious); my Twitter audience’s most liked compositions are not very strong according to my own aesthetic. Perhaps this means I should not provide any initial aesthetic model in the ZF; it would just learn from it’s audience and start with totally uniformly distributed compositions without initial filtering.
The next stage is to integrate this new code into the ZF prototype so that compositions are uploaded once they receive > 50 frames of attention. At the end of a day (when the camera detects darkness) the ZF will download the likes and retweets from Twitter. I’m also thinking through how to store this data (post IDs and feature vectors for compositions) over the long-term and also for the ML process.