MAVA Ontology

A collection of videos that share a common context or purpose.

A media resource consisting of a sequence of images creating the illusion of movement.

• has member ^op
• has video ^op

A .mediapkg archive: a ZIP containing a manifest.json and Parquet annotation files for one or more videos. The unit of interchange between tools.

A series of interval-based annotations for a video. Each row covers a time span with a start and end time. Maps to an interval Parquet file in .mediapkg.

A series of interval-based annotations with multiple labels per interval for a video. Each row covers a time span with a start and end time. Maps to an Parquet file in .mediapkg with a list column.

A single annotation: a time interval with associated value(s). Maps to one row in an annotation Parquet file.

• belongs to annotation series ^op
• start time ^dp
• end time ^dp

• Annotation Segment ^c
• Annotation List Segment ^c

A single-valued annotation: one time interval with one string value.

A multi-valued annotation: one time interval with multiple string values.

A dense time-series of numeric observations sampled at regular intervals. The dimensions (what is measured) are declared via mava:hasDimension. Maps to a time-series Parquet file in .mediapkg.

• has dimension ^op
• sampling interval ^dp

A set of numeric values recorded at a single point in time. One row in a time-series Parquet file, with one value per dimension.

• belongs to observation series ^op
• numeric value ^dp

A single measured quantity within an ObservationSeries, corresponding to one column in the Parquet file. Examples: 'angry', 'explosion_score', 'interior_exterior'.

• dimension name ^dp
• dimension description ^dp
• value range ^dp

A series of spatial detections over time, in long format (one row per detection). Maps to a region Parquet file in .mediapkg. Sibling of ObservationSeries.

A single spatial detection at a point in time: a bounding box with a detection score and identity. One row in a RegionSeries Parquet file. Sibling of ObservationPoint.

• belongs to region series ^op
• x ^dp
• y ^dp
• width ^dp
• height ^dp
• detection score ^dp
• cluster id ^dp
• identity label ^dp

Relates a VideoCorpus or Video to the MediaPackage containing its annotations.

Relates an ObservationSeries to the dimensions it measures.

Relates a VideoCorpus to the videos it contains.

Relates a MediaPackage to the videos it contains annotations for.

Relates a Video to one of its annotation, observation, or region series.

Relates an Annotation back to its parent AnnotationSeries.

Relates an ObservationPoint back to its parent ObservationSeries.

Relates a RegionDetection back to its parent RegionSeries.

Containment parent of a series (the tree it lives in). A series has at most one parent and the parent graph is acyclic.

Provenance: a source series this series was computed from. May have several (1..n) sources.

The name of a dimension, matching the column name in the Parquet file.

Human-readable description of what this dimension measures.

Description of the value range, e.g. '[0,1]', 'boolean', 'categorical'.

Human-readable description of what a series contains.

For ObservationSeries: the interval between samples in seconds, e.g. 0.5.

Time of an ObservationPoint or RegionDetection in seconds from the start of the video. Maps to the 'start_seconds' column in time-series and region Parquet files.

Start time of an AnnotationSegment in seconds from the start of the video.

End time of an AnnotationSegment in seconds from the start of the video. Primary field — duration is derivable as (endTime - startTime).

A numeric value at an ObservationPoint. In Parquet, each dimension has its own column — the column name identifies the dimension.

A string annotation value. Used for transcript text, shot labels, face emotion labels. Maps to 'annotations' column in interval Parquet files.

A list of string values for an AnnotationListSegment. Maps to the 'annotations' column (Parquet LIST type) in list annotation files.

Left edge of a detection's bounding box, normalized to [0,1] of the frame width. Maps to the 'x' column.

Top edge of a detection's bounding box (top-left origin), normalized to [0,1] of the frame height. Maps to the 'y' column.

Width of a detection's bounding box, normalized to [0,1] of the frame width. Maps to the 'w' column.

Height of a detection's bounding box, normalized to [0,1] of the frame height. Maps to the 'h' column.

Detection confidence in [0,1]. Maps to the 'det_score' column.

Machine cluster a detection belongs to (a stable integer). Not necessarily 1:1 with identityLabel — several clusters may map to one label. Maps to the 'cluster_id' column.

Human-provided identity label for a detection (e.g. a person's name). Nullable and incomplete; not unique and not 1:1 with clusterId. Maps to the 'label' column.

How a derived series was computed, e.g. 'argmax', 'cluster_to_scalar', 'aggregate_scalar'. Required whenever derivedFrom is present.