1 · Original Sintel video
Full motion clip. This video provides the frames used by both LK variants below.
original MP4
If playback is blocked: open / download original MP4
Why Multi-Scale Optical Flow?
This version uses the full Sintel MP4 clip. The original video, adjacent-frame ground-truth flow video, and the two Lucas–Kanade estimates run together. Increase the temporal gap to make the compared frames farther apart: single-scale LK should lose correspondences first, while pyramidal LK can recover more of them.
2 · Ground-truth dense flow video
Ground-truth dense flow for adjacent Sintel frames, synchronized to the source video.
adjacent-frame GT
GT asset: open / download GT MP4
One comparison, one input: both LK panels use the same pair of frames from the playing MP4. At Temporal gap = 1, the pair is adjacent; larger gaps deliberately create larger displacement. The GT video above is an adjacent-frame reference, so for gap > 1 it is not the exact GT of the wider-gap comparison.
3 · Single-scale Lucas–Kanade
Full resolution only. Cyan arrows are successful estimates; red circles mark points recovered by the pyramid but missed here.
full resolution · zero initialization
tracked points—
median displacement—
4 · Multi-scale / pyramidal LK
Same frames and same points. Green arrows are valid estimates; yellow arrows are correspondences specifically recovered by the pyramid.
4 resolutions · coarse → fine
tracked points—
recovered by pyramid—
5 · Why the pyramid helps
The same physical motion occupies fewer pixels at coarser resolutions, making LK's local linearization easier.
same motion · smaller at coarse scale
motion → motion/2 → motion/4 → motion/8
Loading MP4 assets…
Classroom takeaway: start at Temporal gap = 1, then increase it toward 5–9 frames. As displacement grows, single-scale LK should lose more correspondences. Pyramidal LK can first solve a smaller displacement at a coarse level and then refine it at finer levels.
MP4 assets are from the public MMFlow Sintel demo. The sparse Lucas–Kanade implementation and image pyramid run locally in the browser.