A brief action between frames: reviewing AI video analysis
A short event can fall between sampled frames. Check the scope, the meaning of timestamps and the difference between a missed event and an absent one.
Article language: English

In this article
"There was no such text" is a stronger conclusion than "I did not find it in the frames I analysed." When a detail appears briefly, the difference may come from how the material was sampled. The model may not have misread a visible number at all. Perhaps none of the selected frames contained it.
A review of video analysis should therefore establish which parts actually entered the assessment. The duration of the accepted file does not describe that scope. Before reporting on how well a system understands a video, establish the sampling rules, an identifiable recording version and a criterion for the event. This is particularly useful when reviewing product demonstrations or material for a publication description.
Six frames from five seconds
Let's work through a time-grid example. We are not creating a video or running AI. In a hypothetical recording, a card bearing the number 7 is visible from 2.2 seconds until just before 2.4 seconds. We take frames at 0, 1, 2, 3, 4 and 5 seconds. The number does not appear in that set.
At four samples per second, starting at zero, one frame falls at 2.25 seconds. That is within the period when the card is visible. One sample per second starting at an offset of 0.3 seconds also hits that period, because one sample falls at 2.3 seconds. This is an exercise in presence at points in time, not a character-recognition result.
The example shows why both frequency and the position of the sampling grid matter. It does not establish a universal number of frames needed to analyse every video. Another event could be shorter, occur at a different time or be obscured. Four samples per second do not become a guarantee just because they helped in this one calculation.
Gemini documentation describes a static mode that samples one frame per second and warns that fast motion or scene changes can make details easy to miss. It distinguishes that mode from one that navigates the material dynamically. When reviewing a particular configuration, record the mode used instead of assigning one rule to every video-analysis service.
The number of frames in the source video and the number of samples used for analysis are different values. A recording can show smooth motion during ordinary playback while the model receives a sparser sequence. A person watching the original should not automatically assume the model received the same continuity. It is a property of the whole input-preparation pipeline.
An Aeneia Audit review of AI use can help organise the purpose of the analysis, how the video is supplied and the criteria for accepting the answer. For a briefly visible marking, a question about which frames were sampled may be more useful than immediately comparing broad model rankings. The process owner should know what the result actually establishes.
What counts as an event
"Check the demonstration" is too broad a requirement. A fictional studio preparing instructions for folding a paper package might ask whether a number appears, where a tab sits or in what order folds are made. Those are different tasks. Recognising a state at one moment does not confirm the whole action that led to it.
If the question concerns the number 7, the reference should identify the period when it is legible. If it concerns turning the package, the beginning and end of the action and the visible basis for recognising it need to be defined. Two shots of a similar box do not necessarily establish continuous movement between them. An edit may conceal a different preparation.
Suppose the tab is open at first and closed later. Those states describe a change, but they do not establish which hand performed it or whether an additional tool was used. If the task is to check the instructions, the manner of execution may determine the result. Do not replace that requirement with the easier question of the final appearance. The record can separate two findings: the final closure is visible; the available samples did not cover how it was closed. That is enough for an editor to know which part to watch personally.
Checking continuity of objects and actions in AI video (in Polish) helps identify which features of a demonstration need to be preserved. When analysing existing material, those features become part of the assessment reference. If the box changes shape between shots, the model should describe the material available rather than inventing missing, correct steps in the instructions.
Prepare one version with the event and one without it. In the latter, it should not appear in a caption, file name or comment either, if the assessment concerns visual recognition alone. Otherwise, the answer may use a cue. Record which information channels were available: frames, sound, subtitles or textual instructions.
A human reference also needs an explicit scope. The person preparing the reference answer should watch the entire segment being assessed and pause the relevant part for a brief marking. "Number 7 present" without a location does not resolve a later dispute about another copy of the recording. If a second person cannot confirm the mark, set the case aside for clarification instead of calling the model wrong on the basis of an uncertain label. A reference does not become infallible because its file is named "final".
Move the brief event along the timeline while preserving its meaning. If the result depends on a shift of a fraction of a second, examine the samples. This does not automatically prove a particular cause, but it is a reason to check the scope of the input. The requirement should reflect the actual pace of material the team will use.

Separately check an event that is partly obscured and one that cannot be recognised from the prepared input. "Not enough evidence to determine" may then be the right result. Do not reward a confident "it did not happen" merely for being concise. The reader needs to distinguish a negative finding from an observation that did not cover enough.
Keep the context before and after the event as well. Cutting out only the middle may make text easier to read while removing the basis for determining the order of actions. The rules for cropping and scaling images before a model (in Polish) apply to every sampled frame here. The necessary detail has to survive both the choice of time and the preparation of the image.
A timestamp has a particular precision
A result with many digits looks precise. It does not necessarily identify the first frame of an event. The Amazon Rekognition reference describes Timestamp in milliseconds from the start of the video and says it is not guaranteed to be accurate to the individual frame where a label first appears. That is the stated scope of this particular service response, not a measurement of every model's accuracy.
The review should establish an acceptable time difference for the intended use. A broad indication may be enough to find an approximate place in a long tutorial. Aligning a brief caption with an action needs a different tolerance. The criterion should follow the intended user's task, not the number of digits returned by the system.
Establish what the indication means before calculating a difference. Should the model give the start of visibility, any moment within the event or an interval to watch manually? In our calculation, 2.25 seconds lies within the period when the number is present. It is not the start of that period. Changing this definition during review will change the assessment of the same number. For an interval, compare both ends and its usefulness: pointing to an entire five-minute recording formally includes a brief piece of on-screen text, but does not help the editor find it.
Write down the time unit. MDN describes the VideoFrame timestamp property as a value in microseconds. Another interface may use milliseconds or a readable minutes-and-seconds notation. Moving the number between them without conversion will not preserve its meaning.
In the report, indicate whether the time refers to the original, a clip or a version with changed playback speed. If the clip starts at 2:00 in the original, its fifth second corresponds to 2:05 in the source, provided the speed has not changed and no further cuts were made. Separate the clip's local timeline from the position in the source. With several segments, each indication should retain its own provenance.
A speed change requires a similar distinction. Slowing the material may make a brief action easier to watch manually, but a result expressed against the slowed copy must be mapped correctly back to the original. Label the original and the slowed copy so they cannot be confused when opening a result. Keep a mapping in the review that lets someone open the right moment without guessing.
When the result concerns speech, check preparing audio, channels and silence (in Polish). Sound can be trimmed independently of the image. Video subtitles (in Polish) also need to match the particular version. Correct words placed against the wrong action still mislead the audience.
The report should show the scope of observation
Before attributing an error to the model, check that the task is complete: the video was accepted, the required segments were processed and the answer belongs to that attempt. A failure to fetch a clip is not a finding that "the event did not happen". No answer also needs a distinct label. One shared empty row conceals different causes.
When results are retrieved in batches, check the completeness of a paginated read (in Polish). An event missing from the first page of a response does not support a conclusion about the rest of the recording. Likewise, a retry for one clip should not be counted as a new, separate segment. Link results to a time range and a particular task. The team needs to distinguish five completed segments from five answers about the same segment.
If a list of the frames actually used is available, retain their timestamps and their relationship to the original. If the service does not expose an exact list, state that limit. Do not reconstruct a supposedly observed range from the video duration alone. In a mode that selects portions in response to the question, another run may need its own record of what happened.
When there is an error, distinguish between an event that never reached the samples, one that was present but not recognised, one recognised with the wrong time, and an answer that went beyond the available evidence. These distinctions give useful directions for further work. The available report may not always establish the cause, but the missing evidence can be recorded instead of guessed.
A principle from assessing long AI context (in Polish) still applies: accepting material does not confirm use of every necessary piece of information. Video adds the choice of moments and image representation. A claim that a system supports long videos does not, by itself, confirm every brief action in a recording.
After changing the sampling frequency or analysis mode, repeat the assessment on the same variants of the material. Record the cost and time needed to produce a useful result if they are part of the decision. Denser sampling may expand the input, but it does not by itself establish correct interpretation. A better resource-use report does not justify losing an event required by the task either.
A working description of a demonstration can end with a clearly limited decision: the solution finds the specified type of event under the conditions tested, or a particular segment needs further review. The card bearing the number 7 does not need a style review. It needs a chance to appear in the material being analysed at all.







