Using https://github.com/YassineYousfi/comma10k-baseline plus opencv and ffmpeg.
Seems like the model likes detecting curbs as lane lines and sidewalks as drivable sufrace, I assume the training data isn't adequate yet. Works quite well for cars though.
