Motion vector analysis
Also known as: codec motion vectors, encoder motion data
Reading the movement data a codec already stored while compressing, instead of recomputing motion by comparing decoded frames.
Updated
What it means
Every modern Video codec stores motion while it compresses. Rather than keeping each frame whole, it records that a block of pixels moved by some amount since the last one.
Those vectors are a free description of movement sitting inside the file. Motion vector analysis reads them instead of decoding frames and comparing them, which is a large saving on a long upload.
The appeal is arithmetic rather than clever. Getting a motion signal without fully decoding the video is what makes scanning a whole card feasible rather than aspirational.
How it works in practice
It is used as the first sweep over an upload. Vectors give a motion curve per clip, the peaks mark candidates, and only those candidates are decoded properly for a closer look.
That is the same economy Frame sampling provides on the picture side, applied one level further down, before any pixels are reconstructed at all.
Vectors also expose camera movement as a whole frame drift, which helps separate a panning camera from a moving subject before anything expensive runs. Subtracting that drift leaves the local movement, which is the signal a Highlight detection model actually wants.
What to watch out for
Vectors describe compression, not reality. An encoder chooses whatever predicts cheaply, so a vector can point somewhere nothing moved, and a flat low light frame can produce noise that reads as motion.
Availability also varies by file. An All-Intra (ALL-I) recording stores no inter frame vectors at all, so the trick silently stops working and the pipeline has to fall back on decoding and comparing frames.
Resolution is coarse by design. Vectors describe blocks rather than pixels, which is plenty for finding busy stretches and useless for anything needing precise subject position. Anything that has to know where a rider is in the frame still needs real detection on decoded pixels.
How RawClip handles it
The first sweep over an upload reads the movement data already inside the files, so the hours that contain nothing are identified without decoding them, and closer analysis is spent only where something actually moved.