ZipF / Blog / Ground truth, not raw footage

Data

Ground truth, not raw footage

Most physical-AI panels infer the world from hours of footage. ZipF records it. Ground truth in this stack is the event itself: class, track, surface, weather, place, time, and rarity, written as the world happens.

Video archives already sell kilometers to perception teams. They are valuable. They are also the wrong unit. A lab does not want another ten thousand hours. It wants the rare join, geolocated, labeled, and legal to train on. ZipF's panel is event-native. It sees the fact that made the hour matter.

That difference is the product. A dump is pixels. A ZipF event is a pedestrian in the carriageway, three two-wheelers, rain, night, unmarked edge, rarity band 4, faces and plates already gone. Mix of actors, weather, and surface live in the second record. The buyer class already pays in dollars: ADAS systems $57B in 2026, HD maps $4.9B, automotive training datasets $953M.

Who this is for

World-model and ADAS teams get packs and eval. Map desks get change. Insurers get risk. Cities get surface tickets. The same events feed all four. The contract is the packaging.

Coverage thesis

India first, because the inner road is ordinary there and the highway file is thin. Asian cities that share that road fabric are next. The schema is global. Time is the second axis: the same corridor next week. The event is the product.

Ground truth that only stores the frame is an archive-shaped product. ZipF's ground truth includes the rarity of the join and the change since the last pass. That is what makes the dataset useful to a lab (where did we fail?), to a map (what moved?), and to a planner (is this meter still true).

Questions

Keep reading