Sample presentation time without promising all motion survives

For a verified constant-rate segment, taking every kth source frame corresponds to a fixed interval if timestamps and missing-frame policy support that claim. For variable-rate input, the same ordinal stride can produce unequal time gaps. A timestamp-driven sampler can choose the nearest frame, the latest frame not later than a target, or another declared interpolation rule. These are different policies. Nearest selection can choose future presentation content and therefore is not automatically suitable for a strict online feature cutoff.

Store the target times, chosen source PTS values, timestamp errors, repeated selections, and uncovered gaps. If several target times map to one source frame, the output array has repeated images rather than new temporal evidence. Interpolation can create synthetic frames but does not make them independently observed content. A fixed-size tensor is a useful model interface, yet its shape does not establish that the intended temporal span or event ordering was preserved.

For an ideal continuous signal bandlimited below a known highest frequency, exact uniform samples at a strictly higher rate than twice that frequency support the usual reconstruction argument with ideal assumptions. Real video involves finite exposure, occlusion, compression, camera motion, and potentially non-bandlimited changes. An anti-alias filter before downsampling can reduce frequencies outside the chosen band, while also removing information. It does not restore distinctions already aliased in the captured source.

SlowFast explores different temporal rates for complementary pathways, and TimeSformer studies space-time attention choices. Their task results motivate architecture and sampling experiments rather than a theorem that all fast motion is preserved. Report recognition quality by action frequency, duration, and source frame cadence where relevant. A task may tolerate lost high-frequency detail, or may depend exactly on it. Distinguish empirical task sufficiency from lossless reconstruction of an arbitrary physical scene.

Build a timestamp fixture with irregular intervals and a synthetic signal with known frequencies. Test the selected times independently from the learned model so an indexing error cannot be explained away as model weakness. Then evaluate real task performance across rate and clip-span changes under the same split. If the model accepts timestamps as features, verify their units and handling of repeated frames; supplying an ordinal position still does not tell it the actual gap.

The waves lab provides a known analytic component that can be inspected at reachable times. Reset, set both phase difference 0 degrees and amplitude ratio 0, then predict the value at x=0 after four Steps: −sin(2)≈−0.9093. Reset and set both phase 0,ratio 1 for −2sin(2)≈−1.8186. Step spacing is half a second. This fixed-frequency analytic instrument does not downsample a video, run an anti-alias filter, or demonstrate the separate seven-versus-three-hertz example; it keeps the observed function and sampling times explicit.