I'm starting to upload a bunch of 'v2' videos. The 'v2' montage maker fixes a bug where the audio was out of sync. If you have your audio muted then you shouldn't notice any difference between the original and 'v2' versions of the videos. If you want to help out, please comment on any of the videos and give your feedback. Alternatively, go to the boldwren/eye-tracking project on github and say hi.

I've spent a bunch of time over the past few weeks teaching my computer to spot eye contact in videos.
I have cobbled together a bunch of montages from some popular categories. The pipeline goes like this:
* Get a list of categories
* For each category, get a list of videos
* For each video, get a list of thumbnails [~16 for most videos]
* Download all of the thumbnails, and get the computer to look for eye contact [see graphs in my uploaded images]
* Find the categories that contain the most eye contact.
* For each category, find the videos that contain the most eye contact.
* Download the best videos in each category
* For each video, find the segments that contain eye contact
* Cut all of the eye contact segments from all videos into a single video. 
Details can be found here:
 
Things that I have observed:
* The Interactive category is dominated by MHBHJ, and all other videos in that category look a bit out of place.
* The Bukkake category has a couple of videos of Xev Bellringer, which have a very different pace from the rest of the videos.
* I have been watching the videos on mute, so I have no idea how they sound.
I currently use the first page of "HD only, Most Viewed videos Yearly" for each category. In theory I can make these montages from any page of video results.  I will upload more montages when they're generated, and I will try to re-upload videos for each category when I make improvements to my algorithms.
My current backlog of things to try lives at  . If you have any feedback/suggestions, please leave a comment or message me.