The Former Treatise Have I Made

In this work, we seek to quantify the “common” or “typical” movements involved in oral sex performed on males. To do so, we analyze a dataset containing over 108 hours of pornographic video, annotated at each frame with the position of the lips along the shaft of the penis. We use quantization techniques to discover sixteen distinct motions, and using these motions we design and evaluate a system that procedurally generates realistic movement sequences using deep learning. We quantitatively show that this system is superior to simple Markov Chain techniques.

This seems like classic Garbage In, Garbage Out statistics. The training set is pornography videos intended for their viewers’ visual appreciation, but produced with professional actors on various desensitizing drugs and spliced together with editing. Yet the output model is going to be applied to a “realistic” blowjob simulating device intended for users’ tactile ejaculatory response? They have optimized for something completely different!