The time series of frame probabilities is of course highly autocorrelated, but I never bothered to model that due to lack of interest. From memory, a simple single-frame classifier was >95% accurate (this was quite a few years ago, mind you). In my small-sample and not very rigorous testing, this was higher than human rates. Turns out that on average it's pretty hard to tell an ad from content just by looking at the pixels of a single frame.
I am quite confident that it wouldn't be too hard to build an ad detection ML model that would have near-perfect accuracy. That said, an approach based on algorithmically detecting repeated segments of lengths consistent with ad spots would work just as well, if not better.
P.S. One thing I thought was really interesting was that the classifier -- that was only ever shown a binary label (ad/not-an-ad) -- learnt an embedding that grouped together entire categories of things across TV networks and geographies (studio news, weather, traffic reports etc).