← Back to slide 5But the motion is not noise — it has structure
← All interactives
Motion Fingerprints
Tap a dataset and watch its fingerprint form: 6,000 of its own annotated flow vectors, drawn one by one as streaks (hue = direction, length = magnitude). The pooled signature is a property of the dataset, not of any one image.
a sample image
0 of 6,000 vectors drawn ·
the paper's splat of all its vectors
A fingerprint is the empirical distribution of [position, displacement] over all annotated vectors: the bag-of-flow-vectors. No spatial binning, resolution-invariant, and it needs only annotations, no pixels. The estimate converges by about 50K pooled vectors and holds to 16M. Chapter 3 measures directed distances between these clouds; Chapter 4 bins them into the coverage table. The vectors drawn here are a uniform subsample of the exact arrays the paper pooled.