Angular spacing, in degrees, of the orbit poses the harness measures — see
borderSampleFrames.
WHY A GRID AND NOT JUST THE TWO EXTREMES. The rock's endpoints are the poses
FURTHEST from the framed still, but they are not necessarily the poses of
greatest projected extent: a feature sitting ~90° round from the screen-x axis
reaches its maximum projected |x| at an INTERMEDIATE rock angle and comes back
in by the endpoint, so an endpoints-only check can read zero while the frames in
between visibly crop.
5° is fine enough to be equivalent to measuring all 120 frames. The worst pose
is either an endpoint (sampled exactly) or a stationary point of the projected
extent, and at a stationary point the extent falls off only as 1 − cos δ: with
the true worst pose at most step/2 = 2.5° from a sampled one, we see it to
within 0.1% of its extent. Halving the step would buy 0.03%; measuring every
frame costs 120 in-page PNG decodes per tile to buy the same.
Angular spacing, in degrees, of the orbit poses the harness measures — see borderSampleFrames.
WHY A GRID AND NOT JUST THE TWO EXTREMES. The rock's endpoints are the poses FURTHEST from the framed still, but they are not necessarily the poses of greatest projected extent: a feature sitting ~90° round from the screen-x axis reaches its maximum projected |x| at an INTERMEDIATE rock angle and comes back in by the endpoint, so an endpoints-only check can read zero while the frames in between visibly crop.
5° is fine enough to be equivalent to measuring all 120 frames. The worst pose is either an endpoint (sampled exactly) or a stationary point of the projected extent, and at a stationary point the extent falls off only as 1 − cos δ: with the true worst pose at most
step/2 = 2.5° from a sampled one, we see it to within 0.1% of its extent. Halving the step would buy 0.03%; measuring every frame costs 120 in-page PNG decodes per tile to buy the same.