Data Formats

DIVE Desktop and Web support a number of annotation, configuration, and media-side metadata formats. The annotation and configuration formats below can be uploaded or imported alongside your media and will be automatically parsed.

• DIVE Annotation JSON (default annotation format)
• DIVE Configuration JSON
• KWCOCO Species List (configuration: the classes a dataset may use)
• VIAME CSV
• KPF (KWIVER Packet Format)
• COCO and KWCOCO

Frame metadata sidecars are media files rather than annotation imports. See Frame Metadata Sidecars for their naming, placement, and text-file format.

DIVE Annotation JSON

Info

The current DIVE schema version is v2. Version 2 was introduced in DIVE version 1.9.0. It is backward-compatible with v1.

Files are typically named result_{dataset-name}.json. Their schema is described as follows.

/** AnnotationSchema is the schema of the annotation DIVE JSON file */
interface AnnotationSchema {
  tracks: Record<string, TrackData>;
  groups: Record<string, GroupData>;
  version: 2;
  /**
   * Annotation frame rate when present. Omitted when absent or unusable.
   * Same role as the VIAME CSV `# metadata` `fps` field and COCO `videos[].annotation_fps`.
   */
  fps?: number;
}

interface TrackData {
  id: AnnotationId;
  meta: Record<string, unknown>;
  /**
   * Track-level attributes. Can contain arbitrary key-value pairs.
   * 
   * Reserved attribute names (cannot be created by users):
   * - `userCreated`: Internal flag indicating user creation status.
   */
  attributes: Record<string, unknown>;
  confidencePairs: Array<[string, number]>;
  begin: number;
  end: number;
  features: Array<Feature>;
}

interface GroupData {
  id: AnnotationId;
  meta: Record<string, unknown>;
  attributes: Record<string, unknown>;
  confidencePairs: Array<[string, number]>;
  begin: number;
  end: number;
  /**
   * members describes the track members of a group,
   * including sub-intervals that they are participating in the group.
   */
  members: Record<AnnotationId, {
    ranges: [number, number][];
  }>;
}

interface Feature {
  frame: number;
  flick?: Readonly<number>;
  interpolate?: boolean;
  keyframe?: boolean;
  bounds?: [number, number, number, number]; // [x1, y1, x2, y2] as (left, top), (bottom, right)
  geometry?: GeoJSON.FeatureCollection<GeoJSON.Point | GeoJSON.Polygon | GeoJSON.LineString | GeoJSON.Point>;
  fishLength?: number;
  /**
   * Detection attributes. Can contain arbitrary key-value pairs.
   * 
   * Reserved attribute names (cannot be created by users):
   * - `rotation`: Rotation angle in radians for rotated bounding boxes (counter-clockwise).
   *   When present, the `bounds` field represents an axis-aligned bounding box, and the
   *   actual rotated rectangle is computed by applying this rotation around the bbox center.
   *   Only stored if rotation is significant (|rotation| > 0.001 radians).
   * - `userModified`: Internal flag indicating user modification status.
   */
  attributes?: Record<string, unknown>;
  head?: [number, number];
  tail?: [number, number];
  /**
   * Free-form note text for this detection (frame). Stored as a string
   * array for format compatibility; DIVE typically uses a single entry.
   * Omitted when the detection has no note.
   */
  notes?: string[];
}

Reserved Attribute Names

Reserved Attribute Names

Certain attribute names are reserved by DIVE and cannot be used when creating custom attributes. Attempting to create attributes with these names will result in an error.

Reserved Detection Attributes (stored in Feature.attributes): - rotation: Used to store the rotation angle in radians for rotated bounding boxes. When present, the bounds field represents an axis-aligned bounding box, and the actual rotated rectangle is computed by applying this rotation around the bbox center. Only stored if rotation is significant (|rotation| > 0.001 radians). - userModified: Internal flag used by DIVE to track user modification status.

Reserved Track Attributes (stored in TrackData.attributes): - userCreated: Internal flag used by DIVE to track user creation status.

These reserved names are enforced at both the UI level (when creating attributes) and the API level (when saving attributes). If you need to use similar names, consider alternatives like rotationAngle, isUserModified, or isUserCreated.

The full source TrackData definition can be found here as a TypeScript interface.

Annotation frame rate (fps)

Optional top-level fps carries the dataset annotation frame rate — the same value VIAME CSV writes in the # metadata header and COCO/KWCOCO records on videos[].annotation_fps.

{
  "version": 2,
  "fps": 5,
  "tracks": {},
  "groups": {}
}

• On export, DIVE writes a usable dataset fps (finite number greater than zero).
• On import, that value is restored into dataset metadata. Unusable values (0, negative, non-numeric, inf/nan) are ignored. A file with no fps leaves the dataset rate unchanged.
• Older v1 track-map files have no place for fps; only the v2 document form carries it.

Example JSON File

This is a relatively simple example, and many optional fields are not included.

{
  "version": 2,
  "fps": 5,

  "tracks": {
    // Track 1 is a true multi-frame track
    "1": {
      "id": 1,
      "meta": {},
      "attributes": {},
      "confidencePairs": [["fish", 0.87], ["rock", 0.22]],
      "features": [
        { "frame": 0, "bounds": [0, 0, 10, 10], "interpolate": true },
        { "frame": 3, "bounds": [10, 10, 20, 20], "attributes": { "rotation": 0.785 } },
      ],
      "begin": 0,
      "end": 2,
    },
    // Track 2 is a simple single-frame bounding box detection
    "2": {
      "id": 2,
      "meta": {},
      "attributes": {},
      "confidencePairs": [["scallop", 0.67]],
      "features": [
        { "frame": 3, "bounds": [10, 10, 20, 20] },
      ],
      "begin": 3,
      "end": 3,
    },
  },

  "groups": {
    "1": {
      "id": 1,
      "meta": {},
      "attributes": {},
      "confidencePairs": [["underwater-stuff", 1.0]],
      "members": {
        // The fish is a group member on frame 0, 1, and 3.
        // The scallop is only a group member at frame 3.
        "1": { "ranges": [[0, 1], [3, 3]] },
        "2": { "ranges": [[3, 3]] },
      },
      "begin": 0,
      "end": 2,
    }
  }
}

DIVE Configuration JSON

This information provides the specification for an individual dataset. It consists of the following.

• Allowed types (or labels) and their appearances are defined by customTypeStyling and customGroupStyling.
• These fields are per-dataset styles in the portable Configuration File.
• Separately, when Type color scope is Shared, DIVE also keeps a cross-dataset style store (Desktop: global_style_settings.json in the data storage path; Web: browser localStorage). That shared store is not part of this Configuration JSON export format.
• Preset confidence filters for those types are defined in confidenceFilters
• Track and Detection attribute specifications are defined in attributes
• Free-form, dataset-level metadata (cruise id, station id, location, …) is stored in datasetInfo as a key/value object.
• Edited from the Dataset Info panel.
• Included in DIVE Configuration JSON as datasetInfo.
• Included in VIAME CSV and COCO / KWCOCO export, and restored on import.
• Annotation frame rate is stored as dataset fps.
• Included in DIVE Annotation JSON as top-level fps.
• Included in VIAME CSV as the # metadata fps field.
• Included in COCO / KWCOCO as videos[].annotation_fps for video datasets.
• A track type hierarchy is stored in typeHierarchy as a child-type to immediate-parent-type map.

For example, this configuration makes fish a heading-only parent (it does not need to be an explicit configured type or appear in a track):

{
  "typeHierarchy": {
    "shark": "fish",
    "great white shark": "shark"
  }
}

A type hierarchy is a single-parent forest. Child and parent names must be non-empty strings, self-edges and cycles are invalid, and each child can have only one immediate parent. Names are preserved exactly; whitespace is used only to determine whether a name is empty.

A missing typeHierarchy leaves the saved hierarchy unchanged. On overwrite import or direct save, null and {} delete it, while a non-empty map replaces it completely. Additive import follows JSON merge semantics: null deletes the hierarchy, {} makes no change, and a non-empty map adds edges to the existing hierarchy. Identical edges coalesce; a different parent for an existing child or a cycle rejects the whole configuration without changing it. Invalid saves and imports report Type hierarchy is invalid: {reason}. No configuration was changed.

DIVE Configuration JSON exports include a valid non-empty hierarchy and omit an absent or empty one. The config.json embedded in a dataset zip follows the same rules. Invalid stored hierarchy prevents either configuration export and reports Type hierarchy is invalid: {reason}. No configuration file was exported. It prevents KWCOCO export and reports Type hierarchy is invalid: {reason}. No COCO file was exported. Hierarchy is not transported by DIVE Annotation JSON, VIAME CSV, KPF, NIST, or labels.txt. KWCOCO transports it through category supercategory fields as described in COCO and KWCOCO, including a hierarchy classification example.

When importing a DIVE Configuration JSON with datasetInfo, Overwrite import (the default) replaces the existing datasetInfo block; an additive import merges it per-key (imported values win). A configuration file with no datasetInfo entry leaves existing dataset metadata untouched.

The full DatasetMetaMutable definition can be found here.

interface DatasetMetaMutable {
  version: number;
  typeHierarchy?: Record<string, string> | null;
  customTypeStyling?: Record<string, CustomStyle>;
  customGroupStyling?: Record<string, CustomStyle>;
  confidenceFilters?: Record<string, number>;
  imageEnhancements?: ImageEnhancements;
  attributes?: Readonly<Record<string, Attribute>>;
  datasetInfo?: Record<string, unknown>;
}

imageEnhancements stores viewer display settings (brightness, contrast, saturation, sharpen, and optional percentile stretch bounds). See Image Adjustments for platform support of high bit-depth stretch.

Media frame metadata

Each frame in an image-sequence or multicam dataset may carry a timestamp field (epoch seconds) parsed from the filename at load time. When every frame on every camera in a multicam dataset has a timestamp, DIVE builds a global aligned timeline for playback. See Aligned playback and timestamps.

interface FrameImage {
  url: string;
  filename: string;
  id?: string; // large-image item id (web tiled TIFF)
  timestamp?: number; // capture time in epoch seconds, when parseable from filename
}

KWCOCO Species List

A species list pre-loads the classes readers pick from, so a dataset opens with the whole list already in the Type List instead of being typed in one video at a time.

The format is a KWCOCO categories block and nothing else — no images, no annotations:

{
  "categories": [
    { "id": 1, "name": "Sebastes" },
    { "id": 2, "name": "Sebastes melanops", "supercategory": "Sebastes" },
    { "id": 3, "name": "Sebastes flavidus",  "supercategory": "Sebastes" }
  ]
}

This is the only species-list format DIVE reads. A file that carries media or annotations is an ordinary COCO / KWCOCO annotation import, even when its annotation list is empty.

• Each name becomes a type the dataset declares, listed in the Type List under Show Empty and selectable in locked mode.
• supercategory — or a one-element parents array — becomes a type hierarchy edge, read exactly as it is for an annotation import.
• Nameless category slots are skipped, with the same warning a COCO import reports.
• A repeated name fails the import. Two slots claiming the same class may disagree about its parent, and the category block of an annotation file would only drop its hierarchy; a species list is imported for its classes, so it is refused instead.
• A species list never creates, changes, or removes annotations.

Importing a species list

• In the viewer, use Import and choose the file, on Web and on Desktop.
• At upload, put it in the Species List field on the Web upload page or the Desktop import dialog. On Web it may be uploaded together with an annotation file and a DIVE Configuration JSON; one species list per dataset.
• Beside the media, name it to end in species.json (for example, rockfish.species.json) and it is picked up automatically when the folder is imported. For a Desktop multicamera import it is looked for in the folder the cameras share, then beside each camera; if the cameras carry different lists none is applied and the import dialog warns, so you can pick one in its Species List field.

Overwrite and additive imports

The import dialog's Overwrite checkbox decides how a list meets what the dataset already declares:

• Overwrite (the default) makes the file the whole declaration. Types it omits stop being declared and the hierarchy is replaced; a list with no supercategory clears the stored hierarchy. Styles are kept for the types the file names.
• Additive adds the file's species and hierarchy edges and keeps everything already declared. A list with no supercategory leaves the stored hierarchy alone.

Neither mode can orphan annotations: a type a track actually uses is listed from that track's confidence pairs whether or not it is declared. Removing it from the declaration only drops its saved color, which falls back to the default palette.

A list whose hierarchy cannot be applied — a cycle, a self-edge, or a child given two different parents — fails the import with Type hierarchy is invalid: {reason}. No configuration was changed. rather than importing a flat list, and a list that repeats a name fails with Species list repeats category names: {names}. No configuration was changed. Unlike the category block of an annotation file, which degrades to a warning, a species list is imported for its classes.

For a multicamera dataset the declared types and the hierarchy are stored on the parent, so a species list imported against one camera updates the whole dataset.

VIAME CSV

Read the VIAME CSV Specification.

Warning

VIAME CSV is the format that DIVE exports to. It doesn't support all features of the annotator (like groups) so you may need to use the DIVE Json format. It's easier to work with.

Dataset metadata in the header

DIVE writes a # metadata comment line near the top of the CSV carrying dataset-level values such as fps. When a dataset has Dataset Info custom metadata, the whole datasetInfo object is added to that line as a single nested JSON entry keyed dataset_info:

# metadata,fps: 23.976,"dataset_info: {""gfishsite_id"": ""2024TXN012"", ""year"": ""2024""}", ...

• On import the # metadata line is parsed back into dataset metadata.
• fps and the dataset_info block are restored; other fields (such as exported_by) are ignored.
• Overwrite import (the default) replaces the existing dataset_info block; an additive import merges it per-key (imported values win).
• A CSV with no dataset_info entry leaves existing metadata untouched.
• This is how dataset context, for example a gfishsite_id used to re-link annotations to an external database, travels with the exported annotations without renaming files. See the Dataset Info panel for how to populate it.

VIAME CSV polygons and length

DIVE extends standard VIAME CSV with additional geometry and measurement fields:

Polygons — one or more (poly) columns per row, each followed by flat x y coordinate pairs:

0,1.png,0,100,100,500,500,1.0,-1,fish,1.0,(poly) 100 100 200 100 200 200 100 200

Multiple polygons — additional (poly) columns on the same row:

..., (poly) 100 100 200 100 200 200 100 200, (poly) 300 300 400 300 400 400 300 400

Holes — (hole) columns follow the outer (poly) they belong to:

..., (poly) 100 100 500 100 500 500 100 500, (hole) 200 200 400 200 400 400 200 400

Multiple holes are supported with additional (hole) columns.

Length measurements — the standard VIAME length column (8th numeric field) stores stereo fish-length values. DIVE also reads and writes a length entry in detection attributes; on export, the resolved value is written to both the column and attributes when present. Interactive stereo in DIVE Desktop populates these values during annotation.

VIAME CSV notes

DIVE stores a free-form note on each detection (Feature.notes). In VIAME CSV it is written as a (note) column on the detection row:

0,1.png,0,100,100,500,500,1.0,-1,fish,1.0,(note) primary observation

• The (note) column is followed by the note text (whitespace after (note) is trimmed on import).
• Notes are per-detection (CSV row), not track-level. There is no track-scoped note token.
• On import, (note) text is stored in Feature.notes (typically a single-element string[]).
• On export, the note is emitted last among the token columns so DIVE Desktop and Web produce identical row order.
• If a row contains more than one (note) column, all values are imported into Feature.notes and preserved on re-export, but the DIVE UI edits a single combined note string.

KWIVER Packet Format (KPF)

DIVE supports MEVA KPF

• Read the KPF Specification
• See example data in meva-data-repo

Info

KPF is typically broken into 3 files, but DIVE only supports annotations being loaded as a single file. However, the 3-file breakdown is just convention and KPF can be loaded from a single combined file.

# Example: create a sinlge KPF yaml annotation file for use in DIVE
cat 2018-03-07.11-05-07.11-10-07.school.G339.*.yml > combined.yml

COCO and KWCOCO

DIVE Web and Desktop can import and export COCO for a single dataset at a time (an image-sequence dataset or a single video dataset). KWCOCO-compatible files are also accepted on import.

When Checked Types Only is enabled, export matches the checked names against each track's raw stored confidence pairs and removes nonmatching pairs from the exported vector. It does not replace that evidence with the hierarchy-resolved type currently displayed in the viewer.

• Read the COCO Specification
• Read the KWCOCO Specification

Export Notes

• For image-sequence datasets, exported images[].file_name uses dataset image filenames.
• For video datasets, DIVE exports per-frame synthetic names (for example, frame_000123.jpg) because base COCO does not define a canonical video container field.

DIVE KWCOCO Classification Profile

DIVE Web and Desktop use the same KWCOCO profile for hierarchy and complete confidence vectors:

• categories contains every type in an exported confidence vector and every child or parent in the dataset hierarchy. A child's immediate parent is written as supercategory. On import, that field is the parent edge; a one-element parents list is used only when supercategory is absent.
• Every annotation retains standard category_id and score fields for the highest-confidence exported pair, so readers that ignore KWCOCO extensions still receive a primary category.
• Every annotation also has a dense prob array aligned by position with the document's complete categories array.
• dive_confidence_pairs stores the track's ordered sparse vector exactly. This preserves the difference between a missing pair and a pair explicitly scored 0, which a dense prob array cannot express. The extension is listed in info.dive_extensions and takes precedence when a DIVE-authored file is imported again. A present but malformed extension produces one import warning and falls back to a valid prob vector or the primary category and score.

For an external KWCOCO file without dive_confidence_pairs, DIVE maps prob by the original category-array order, including unnamed positional slots. It accepts finite numeric values, clamps them to [0, 1], keeps the ten highest entries above 0.001, and falls back to category_id plus score when the vector length is wrong or duplicate category names make the mapping ambiguous. For a track whose annotations contain different vectors, the annotation at the highest frame index wins; the greater annotation ID wins a same-frame tie, independent of file order.

Categories with missing names, duplicate names, multiple parents, invalid edges, or cycles produce an import warning. Usable annotations are still imported. In a multicamera import, the first valid camera hierarchy in configured camera order becomes the parent dataset hierarchy. Matching later hierarchies coalesce; conflicting later hierarchies are skipped with a warning. Camera datasets do not retain separate hierarchy copies.

See Example KWCOCO file with hierarchy classifications for a complete document that round-trips a multi-level hierarchy and an exact confidence vector.

DIVE COCO Attribute Extensions

COCO does not define standard fields for arbitrary track or detection attributes or free-form notes. To preserve DIVE attributes and notes during COCO export/import, DIVE uses extension fields on each COCO annotation object:

• dive_detection_attributes: Detection/frame-level attributes (maps to Feature.attributes)
• dive_track_attributes: Track-level attributes (maps to Track.attributes)
• dive_notes: Per-detection note (maps to Feature.notes)

These extension keys are declared in the COCO info object as:

• info.dive_extensions = ["dive_detection_attributes", "dive_track_attributes", "dive_notes", "dive_confidence_pairs"]

Dataset-level metadata (datasetInfo)

The dataset's free-form Dataset Info metadata (e.g. gfishsite_id, cruise, station) is written to the COCO info block under a single dive_dataset_info key and advertised in info.dive_extensions:

• info.dive_dataset_info = { "gfishsite_id": "2024TXN012", "year": "2024", ... }

Annotation frame rate (videos[].annotation_fps)

Neither MS-COCO nor KWCOCO define a frame-rate field. On import, DIVE reads the annotation FPS the same way VIAME writes it: a positive numeric fps on an entry in the top-level videos table (the COCO counterpart of the VIAME CSV # metadata fps header). Image-sequence documents typically omit videos and carry no rate.

{
  "videos": [
    { "id": 1, "name": "clip", "fps": 5 }
  ],
  "images": [
    { "id": 1, "file_name": "frame_000000.jpg", "frame_index": 0, "video_id": 1 }
  ]
}

• A usable value (finite number greater than zero) is restored into dataset metadata as fps. Unusable values (0, negative, non-numeric, inf/nan) are ignored.
• When multiple video entries are present, the first usable fps wins.
• On export of a video dataset, DIVE writes a one-entry videos table with the annotation FPS and sets images[].video_id. Image-sequence exports omit videos so re-import does not treat them as video.

Extension Field Details

The DIVE extension fields are JSON objects with user-defined key/value pairs. Values are typically strings, numbers, or booleans.

• annotation.dive_detection_attributes
• Scope: one COCO annotation (one frame-level detection)
• DIVE mapping: Track.features[i].attributes
• annotation.dive_track_attributes
• Scope: logical track identity across frames (track_id)
• DIVE mapping: Track.attributes
• annotation.dive_notes
• Scope: one COCO annotation (one frame-level detection)
• Type: string[] (typically one entry; a single non-empty string is also accepted on import)
• DIVE mapping: Track.features[i].notes
• Legacy alias: on import, if dive_notes is absent, DIVE also reads notes

When importing, DIVE merges any keys in the attribute objects into the target detection/track attribute dictionaries. If the same key appears in multiple annotations belonging to the same track, later imported entries may overwrite earlier values for that track-level key. The note is attached to the feature for that annotation only.

Round-Trip Behavior

For COCO files produced by DIVE:

• DIVE writes info.dive_extensions to advertise the extension keys used.
• DIVE writes dive_detection_attributes and dive_track_attributes on each annotation when attributes are present.
• DIVE writes dive_notes on each annotation when that feature has a note.
• DIVE writes category-aligned prob plus exact dive_confidence_pairs on each annotation.
• Re-importing that file into DIVE preserves hierarchy edges, track IDs, complete confidence vectors, attributes, and notes.
• For video datasets, DIVE also writes videos[].annotation_fps (and images[].video_id) so annotation FPS round-trips. Image-sequence exports omit videos. See Annotation frame rate (videos[].annotation_fps).

For COCO files not produced by DIVE:

• DIVE still imports standard COCO fields (bbox, optional polygon segmentation, optional keypoints), and reads extension fields when present.

Supported / Unsupported COCO Features

• Supported:
• Bounding boxes (bbox)
• Polygon segmentations in list format (segmentation: [[x1, y1, ...]]); if bbox is omitted, DIVE derives it from the polygon's axis-aligned bounds
• Head/tail keypoints from category keypoint labels
• Partially supported:
• COCO has no direct equivalent for DIVE groups, so groups are not represented in COCO export.
• Partially supported:
• Run-length encoded segmentations (RLE): bounding boxes and other fields import, but masks are skipped and a warning is shown.

Example COCO Annotation with DIVE Extensions

{
  "info": {
    "description": "DIVE export for my-dataset",
    "dive_extensions": ["dive_detection_attributes", "dive_track_attributes", "dive_notes", "dive_confidence_pairs"]
  },
  "images": [
    { "id": 1, "file_name": "frame_000000.jpg", "frame_index": 0 }
  ],
  "categories": [
    { "id": 1, "name": "fish", "keypoints": ["head", "tail"] },
    { "id": 2, "name": "shark", "supercategory": "fish" },
    { "id": 3, "name": "crab" }
  ],
  "annotations": [
    {
      "id": 1,
      "image_id": 1,
      "category_id": 2,
      "bbox": [100, 200, 50, 80],
      "score": 0.97,
      "prob": [0.03, 0.97, 0],
      "dive_confidence_pairs": [["shark", 0.97], ["fish", 0.03]],
      "track_id": 42,
      "dive_detection_attributes": {
        "visibility": "poor",
        "occluded": true
      },
      "dive_track_attributes": {
        "reviewed": true,
        "source": "analyst"
      },
      "dive_notes": ["primary observation"]
    },
    {
      "id": 2,
      "image_id": 1,
      "category_id": 3,
      "bbox": [320, 140, 120, 90],
      "score": 0.91,
      "prob": [0, 0, 0.91],
      "dive_confidence_pairs": [["crab", 0.91]],
      "track_id": 77,
      "segmentation": [
        [320, 140, 360, 130, 430, 170, 440, 220, 360, 230, 325, 200]
      ],
      "keypoints": [350, 150, 2, 410, 210, 2],
      "num_keypoints": 2,
      "dive_detection_attributes": {
        "visibility": "clear"
      },
      "dive_track_attributes": {
        "species_confidence_note": "manual QA"
      }
    }
  ]
}

Example KWCOCO file with hierarchy classifications

A DIVE type hierarchy is a child-to-parent map. This configuration:

{
  "typeHierarchy": {
    "shark": "fish",
    "great white shark": "shark",
    "ray": "fish"
  }
}

is the forest fish → shark → great white shark plus unused sibling ray. KWCOCO stores each immediate parent on the child category as supercategory. DIVE also writes every hierarchy member into categories, including heading-only parents (fish) and unused children (ray).

The annotation below scores great white shark highest, keeps ancestor shark at an explicit 0, and scores unrelated rock. Dense prob is aligned with categories order; missing pairs become 0 there. Sparse dive_confidence_pairs is the source of truth: shark scored 0 is kept, while fish and ray are absent rather than zero.

{
  "info": {
    "description": "DIVE export for my-dataset",
    "dive_extensions": ["dive_confidence_pairs"]
  },
  "images": [
    { "id": 1, "file_name": "frame_000000.jpg", "frame_index": 0 }
  ],
  "categories": [
    { "id": 1, "name": "shark", "supercategory": "fish" },
    { "id": 2, "name": "great white shark", "supercategory": "shark" },
    { "id": 3, "name": "rock" },
    { "id": 4, "name": "ray", "supercategory": "fish" },
    { "id": 5, "name": "fish" }
  ],
  "annotations": [
    {
      "id": 1,
      "image_id": 1,
      "category_id": 2,
      "bbox": [100, 200, 50, 80],
      "score": 0.91,
      "prob": [0, 0.91, 0.22, 0, 0],
      "dive_confidence_pairs": [
        ["shark", 0],
        ["great white shark", 0.91],
        ["rock", 0.22]
      ],
      "track_id": 42
    }
  ]
}

On import, that document restores:

{
  "typeHierarchy": {
    "shark": "fish",
    "great white shark": "shark",
    "ray": "fish"
  }
}

and the track confidence vector [["shark", 0], ["great white shark", 0.91], ["rock", 0.22]].

External KWCOCO files may omit supercategory and use a one-element parents list instead. DIVE treats that as the same parent edge; supercategory wins when both are present.

{ "id": 1, "name": "shark", "parents": ["fish"] }

Multi-point head/tail centerlines

Centerlines reuse named keypoints and the existing HeadTails GeoJSON LineString. An example optional section in a VIAME CSV row is:

(kp) head 100.25 120,(kp) spine_001 140 105.5,(kp) spine_002 180 115,(kp) tail 210 140

Point names are ordered head, numerically sorted spine_N, then tail. Writers preserve subpixel coordinates. The editor assigns zero-padded sequential names; indices may be renumbered after editing and are not cross-camera IDs.

DIVE JSON stores the ordered coordinates in a feature with properties.key = "HeadTails" and geometry.type = "LineString", alongside named Point features. When a line is present its coordinates are authoritative; its point markers are regenerated on import/edit/export to prevent stale vertices. Line-only JSON can therefore be exported to CSV without losing interior points. Two-point head/tail files remain valid. Older DIVE versions may drop the new interior points; use the updated readers and writers for round trips.

COCO import/export supports these centerlines in both desktop and web DIVE. Exports use category keypoints labels (head, spine_001, …, tail), 1-based skeleton edges, and annotation [x, y, visibility] triples plus num_keypoints. A shared label list covers curves with different numbers of vertices; absent slots are [0, 0, 0] and do not become vertices on import. Coordinates retain subpixel precision. The edited HeadTails LineString is authoritative on export, including when it has no separate point markers.

DIVE also imports VIAME's KWCOCO named-point lists, resolving either keypoint_category names or keypoint_category_id through keypoint_categories. Both formats reconstruct the editable line from head, numerically ordered spine points, and tail. Without both endpoints, points are retained without creating a complete line. Other named keypoints are retained separately. These fields store the sampled polyline, not spline coefficients or physical stereo correspondences.