This is very cool. It looks from the videos like the next step for them is to provide some sort of temporal stability so that detected objects don't get temporarily forgotten across frames and so the bounds expand and contract smoothly. It's obvious that the detection is being run frame-at-a-time.

I also wonder to what extent merging the detection with underlying P-frame information from the video codecs would help. Knowing that a segment of video just moved to the left would mean the detected object could be moved to the left by the same amount, even if it was passing behind another object. Calculating the movement vectors independently seems silly if you can get that data from the underlying video codec itself.