# A Bounding Box Is Not a Census

_Published 2026-09-12._

The steppe is large enough to make a delayed answer feel like no answer at all. A ground count can be careful and still miss the night movement, the winter haze, or the part of the route no team could reach.

That is the problem SaigaVision starts with. The goal is not to produce a video with boxes drawn around animals. The goal is to give conservation teams a record of what was observed, where it was observed, and how an identity moved through the pass.

## Detection is an observation

An aerial pass produces two different kinds of input: optical frames and thermal frames. Neither is a universal answer. Optical imagery is useful in conditions where thermal contrast is weak; thermal imagery is useful when darkness or visibility makes the optical channel unreliable.

The first useful output is a detection. It has location, confidence, and time. That is already better than a screenshot, but it is not yet a count.

## Tracking is where the meaning begins

A box on frame 41 and a box on frame 42 might be the same animal. Or they might not. A conservation record needs an identity over time, with the uncertainty kept visible. When a track is lost, the safe choice is to close it and start a new one rather than merge two animals because the story would be cleaner.

The [SaigaVision case study](/work/saiga-vision) keeps the boundary clear: capture, inference, tracking, movement records, and conservation analytics are in scope. Flight control, airframe design, and enforcement action are separate systems.

Videos are valuable evidence, but they are difficult to query. Records make the survey useful after the aircraft lands. A team can compare passes and inspect low-confidence events instead of scrubbing through hours of footage hoping to notice a pattern.