Video Dataset Ontology

FLOW ALGEBRA — VIDEO DATASET ONTOLOGY
Version: 1.0 | Date: 2026-10-29 | Encoding: UTF-8
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/video-dataset-ontology.html
Authors: Kireth Research; companion formalisation by Sky — Recursive AI
Status: Independently constructed companion model; source page content not verified.
Citation: Cite the source URL and this companion formalisation when reused.

PURPOSE
Represent video datasets as versioned collections of temporally ordered visual and optional audio observations, with annotations, provenance, partitions, quality checks and reproducible model-training pathways.

STATE MODEL
S_t = (I_t, R_t, B_t, K_t)
I_t: current video samples, frame streams, timestamps, audio tracks and metadata.
R_t: entity and temporal relationships, alignment and annotation rules.
B_t: constraints on privacy, rights, formats, time, quality and data splits.
K_t: dataset ontology, provenance, schemas and task knowledge.
E_t = expand(I_t, R_t)
V_t = {x in E_t | valid(x, B_t)}
x*_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
I_(t+1) = update(I_t, x*_t, observations_t)
The argmax is an optional ranking policy, not a claim that dataset validity is probabilistic.

ENTITY TYPES
VideoDataset(id, version, licence, provenance, schema)
VideoSample(id, uri, duration, frame_rate, resolution, codec)
Frame(sample_id, index, timestamp, image)
AudioTrack(sample_id, timebase, channels, sample_rate)
Annotation(id, sample_id, task, target, label, interval, annotator)
TemporalSegment(sample_id, start_time, end_time)
ObjectTrack(sample_id, object_id, observations)
Split(dataset_id, partition, sample_ids)
CaptureDevice(id, calibration, metadata)
QualityRecord(sample_id, checks, result)
ConsentRights(sample_id, permissions, restrictions)

RELATIONSHIPS
VideoDataset contains VideoSample.
VideoSample consists_of ordered Frame observations.
VideoSample may_have AudioTrack.
Annotation describes VideoSample, Frame, TemporalSegment or ObjectTrack.
ObjectTrack links observations of a hypothesised object across time.
VideoSample captured_by CaptureDevice.
VideoSample governed_by ConsentRights.
Split assigns samples to training, validation or test partitions.
QualityRecord evaluates VideoSample.

FORMAL OPERATIONS
For sample v, frames F(v) = [(f_i,t_i)]_(i=1..n), with t_(i+1) >= t_i.
Duration D(v) >= 0; 0 <= t_i <= D(v) when timestamps use the sample origin.
Nominal constant-rate frame count n approximately D(v)*fps(v); variable-frame-rate video requires timestamps rather than this approximation.
Temporal segment [a,b): 0 <= a < b <= D(v).
Aligned audio A(t) and video F(t) share a documented synchronisation mapping phi(t).
Preprocessing P(v) = decode -> timestamp -> sample/resize -> normalise -> validate.
Dataset D = union of valid samples after duplicate and rights checks.
Partition D = D_train disjoint_union D_val disjoint_union D_test, when a fixed three-way split is used.
Group-disjoint constraint: no capture-session or subject group may span partitions when independence at that group level is required.
Supervised objective L(theta) = (1/N) * sum_i loss(model_theta(P(v_i)), y_i).
Temporal classification loss may aggregate frame or segment predictions using a declared operator.
Tracking quality and detection metrics depend on task-specific definitions; do not conflate them.

INFERENCE RULES
IF frame timestamps are not monotonic THEN flag temporal integrity error.
IF required annotations are absent THEN exclude from supervised labelled subset, not necessarily from self-supervised training.
IF identical or near-identical capture groups cross evaluation partitions THEN flag possible leakage.
IF audio/video synchronisation drift exceeds configured tolerance THEN flag alignment issue.
IF consent or licence forbids a use THEN block that use regardless of technical validity.
IF frame rate is variable THEN derive temporal features from timestamps, not index alone.
IF annotation confidence is low THEN mark uncertain and request review; do not silently promote to ground truth.
IF sample fails integrity validation THEN quarantine until reviewed.

MINI RECURSION ENGINE
input -> inspect metadata -> decode -> check timestamps -> align modalities
-> validate rights -> verify labels -> check duplicates -> assign split
-> record provenance -> publish versioned manifest -> audit -> repeat on changes.
Stop when all required constraints pass or a blocking issue is reported.

TRI-MAPPING
Entity: VideoSample | Relation: contains | Operation: decode and sample frames.
Entity: TemporalSegment | Relation: part_of VideoSample | Operation: segment timeline.
Entity: Annotation | Relation: describes | Operation: train or evaluate.
Entity: ObjectTrack | Relation: links frames | Operation: temporal association.
Entity: Split | Relation: partitions dataset | Operation: leakage-controlled evaluation.
Entity: QualityRecord | Relation: evaluates | Operation: quarantine or approve.

HOW TO READ
Entities define what exists; relationships define what connects; operations define what can be executed. Constraints decide which transitions are valid. Versioned provenance makes outputs auditable.

EXECUTION GUARANTEES (CONDITIONAL)
If inputs are readable, schemas complete and validation succeeds, the pipeline yields a traceable dataset manifest.
Determinism requires fixed preprocessing configuration, seeds, decoding implementation and dataset version.
No guarantee of model accuracy, absence of bias, legal permission or complete annotation follows from ontology conformance alone.

AI SEARCH FLOW
FLOW_ID: VideoDatasetOntology_v1
SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/video-dataset-ontology.html
INTENT: describe, validate, compare and trace video datasets.
ALIASES: video corpus; video training data; temporal visual dataset; annotated video collection.
SEARCH_PATH: identify task -> find samples -> inspect provenance -> inspect temporal structure -> validate rights -> check annotations -> assess splits -> return grounded result.
RETRIEVAL_KEYS: dataset_id, version, sample_id, modality, frame_rate, duration, labels, provenance, licence, split, capture_group.
CONFIDENCE: evidence-based; report missing fields and unresolved constraints.
UPDATE_MODEL: append immutable revision with change log; invalidate dependent assessments when source records change.