Viviannnyang/VIPOSeg-caption-action-temporal-train

VIPOSeg Grounded Captions and Action Temporal Annotations

This dataset extends VIPOSeg with grounded video captions, exact text-span-to-object correspondences, entity coreference groups, and action temporal annotations. Each row represents one TRAIN video and its complete caption, containing multiple entities and actions.

The release includes all 2,801 TRAIN videos in an eight-column annotation table. The manifest records export counts and format conventions. VAL records are not included.

Repository layout

Path	Contents	Rows	Status
train/verification.parquet	TRAIN captions, entity groups, object masks, exact span links, actions, and temporal segments	2,801	Annotation table
train/manifest.json	Export counts, provenance, checksums, and coordinate conventions	—	Export manifest
README.md	Dataset description and annotation format	—	Dataset card

The original TRAIN split is preserved. Rows follow ascending natural video-ID order.

Export statistics

Item	Count
Video/caption rows	2,801
Original stored frame references	66,942
Entity groups	19,299
Object tracklets	19,297
Present object masks	390,501
Unique text-span links	16,352
Actions	6,987
Temporal segments	7,531
Actions with multiple segments	407

Unique span links can refer to several groups and tracklets, for example when a collective phrase identifies multiple people. Counts of groups, spans, and tracks therefore need not be equal.

Parquet schema

Each row contains exactly the following eight columns. Nested JSON values are serialized as strings for portable Parquet and browser handling.

Column	Contents
video_id	Original video identifier
split	train
text	Complete caption
frame_files_json	Ordered relative paths to all original stored 5 FPS frames
groups_json	Entity groups, mentions, exact character spans, grounding metadata, and tracklet references
span_links_json	Exact text spans and their group/tracklet correspondences
tracklets_json	Source object identities and COCO compressed RLE masks aligned with the frame list
actions_json	Action text/spans, type, mentions, subject/object references, confidence, and temporal segments

Entity and tracklet links

Entity groups retain the original identifiers, category, and mention structure. The following fields describe the group and its grounding:

Field	Meaning
group_id	Original numeric entity-group identifier within the video
mark_id	Original source object identifier; null for an explicitly ungrounded group
category	Entity category
mentions	Mention texts, exact character spans, and associated mention metadata
tracklet_ids	Corresponding object-tracklet identifiers; an empty list for explicitly ungrounded groups

A grounded source object has a video-local tracklet identifier. For example:

viposeg-0002

Tracklet identifiers are scoped by the video identifier. They are not globally unique across videos.

Each span link contains these fields:

Field	Meaning
start	Inclusive starting character offset
end	Exclusive ending character offset
text	Exact caption substring at those offsets
group_ids	Entity groups associated with the text span
tracklet_ids	Object tracklets associated with the text span

A shared span can link to multiple objects. Body-part groups can overlap a person's possessive mention in text while retaining a distinct entity identity.

Coreferent mentions are grouped together. The following examples represent different caption contexts:

Example entity group	Mentions assigned to that group
Book	“a book”, first “it”, second “it”
Person	“A man”, “He”, “his”
Right hand	“his right hand”, the “it” referring to that hand

The person and right-hand groups are distinct. The two explicitly ungrounded groups in this release retain their text identity and have no mask tracklet.

Object masks and frame paths

The source object ID is the same ID used by the caption-grounding marks. Only objects referenced by entity groups are included as tracklets. TRAIN category labels use a derived object-to-category mapping; they are not official category annotations.

All 66,942 stored source-frame references are included in their original order. Every tracklet has one mask entry per frame:

Object visibility	Mask entry
Present in the frame	A COCO compressed RLE mask
Absent from the frame	null

The masks are embedded in the Parquet. Source annotation files use this path pattern:

Annotations/<video_id>/<source_frame>.png

Original RGB images are not bundled. Use the original TRAIN directory as the media root. Source-image references use this relative path pattern:

JPEGImages/<video_id>/<filename>.jpg

For example:

JPEGImages/608_eollizcRpgQ/00001098.jpg

Actions and temporal segments

Each action retains its text, character spans, action type, component mentions, subject, objects, and temporal segments. Key fields include:

Field	Meaning
action_id	Numeric action identifier within the video
action_text	Action text
action_span / action_spans	Action character span and component spans
subject / objects	Participants with their individual texts, spans, and group IDs
subject_group_ids / object_group_ids	Aggregated participant group references
segments	Separate temporal intervals using original frame numbers

Multiple objects keep their individual participant records. Discontinuous action text retains its component spans and particle.

The following action has two temporal segments. Its identifying fields are:

Item	Value
Video	0_wHveSGjXyDY
Action	hammers

Only the identifying and temporal fields are shown below:

{
  "action_id": 1,
  "action_text": "hammers",
  "segments": [
    {"start_frame": 217, "end_frame": 238},
    {"start_frame": 259, "end_frame": 262}
  ]
}

Character spans use zero-based, half-open offsets:

[start, end)

Action-frame intervals use original image filename numbers and include both endpoint frames. For example:

Original frame number	Corresponding image filename
217	00000217.jpg

This number is an original filename identifier, not a list position or a sampled-frame label.

Original frame numbers may have gaps. An interval describes the included frames in the ordered source-frame list, not a requirement that every intervening integer have an image. Separate segments are preserved rather than collapsed into a single interval. Action intervals do not expose seconds, 5 FPS sequence positions, or temporal-source fields.

Data sources

This is a derived annotation release based on VIPOSeg media and its associated segmentation annotations. Caption, entity-linking, action, and temporal annotations are additions to the original dataset annotations. Users should consult the original dataset's access, citation, and licensing terms when obtaining or using its media and annotations.