Viviannnyang/VIPOSeg-caption-action-temporal-val-final

VIPOSeg Caption, Action and Temporal Annotations — VAL Final

This release contains all 345 VAL video annotations currently retained in the local VIPOSeg review application at export time. It preserves the current saved caption, entity identities, grounding choices, actions, participants and displayed 5 FPS temporal intervals. All 345 records were confirmed at snapshot time. Review states, notes and confirmation metadata are omitted from the annotation table. TRAIN is not included.

It follows the eight-column format of TRAIN Final. This is a saved snapshot, not a live synchronized service. “Final” is the release name; it does not establish independent semantic correctness or imply a new visual reannotation of every video.

Files and coverage

Path	Contents
val/verification.parquet	345 rows of captions, entities, span links, masks, actions and temporal segments
val/manifest.json	Counts, checksums, reference revision and coordinate conventions
README.md	Dataset description and schema

The original VAL media contains 348 folders. Three records removed from the review application's current annotation pool are excluded. Rows follow ascending numeric video-folder prefix, then full folder name. Images are not bundled.

Statistics

Item	Count
Video/caption rows	345
Stored source-frame references	8020
Entity groups	2480
Object tracklets	2479
Present object masks	49636
Unique text-span links	2062
Actions	919
Temporal segments	1003
Actions with multiple segments	52

Shared entity phrases can refer to multiple objects. Group, tracklet and unique-span counts therefore need not agree.

Parquet schema

Every row has eight string columns. Nested values are serialized JSON arrays, matching TRAIN Final.

Column	Contents
video_id	Original full video-folder identifier
split	Always val; the dataset loading split is validation
text	Current saved complete caption
frame_files_json	Ordered paths to every stored source frame
groups_json	Current entity groups, mentions, character spans, grounding metadata and tracklet references
span_links_json	Exact text spans and their group/tracklet links
tracklets_json	Object identities and COCO compressed RLE masks aligned with source frames
actions_json	Actions, mentions, participants, confidence and temporal segments

Entities and text spans

Groups retain current group_id, mark_id, category, isthing, mentions, confidence/agreement metadata and explicit ungrounded identities. Added tracklet_ids such as viposeg-0002 are local to a video. Ungrounded groups have no mask tracklet.

Character spans are zero-based and half-open: [start,end). Every span must slice to its exact text and lie on complete-word boundaries. A pronoun it cannot refer to the letters inside white, with, sits or knitted. Entity mentions retain one-based, case-sensitive occurrence metadata normalized to round-trip to their authoritative span under overlapping raw substring scanning. Current participant text and saved group assignments are retained; participant positions follow the exact full-word occurrence represented by the current caption/entity layer. Missing derived action fields are constructed from the retained action mentions, without inventing actions or changing their order.

Span links contain start, end, text, group_ids and tracklet_ids. Shared spans may link to several groups. Legitimate overlapping person, possessive and body-part phrases are retained.

Source frames and masks

Relative frame references are JPEGImages/<video_id>/<original_filename>.jpg. Obtain the original VAL media separately and use it as the media root. Each tracklet retains identity, source annotation ID, category, present_frames and exactly one masks entry per stored frame: COCO compressed RLE with size and string counts, or null when absent.

VAL full-sequence masks are taken from Annotations_gt/<video_id>/<original_frame_number>.png, not the sparse Annotations folder. Object IDs are the indexed segmentation IDs used by the current reviewer's entity links. Only currently referenced grounded objects are exported. Every encoded mask was decoded and compared to its original source pixels before release.

Actions and displayed temporal intervals

Each action contains action_id, action_text, action_span, action_spans, action_type, particle, sentence_index, mentions, subject, objects, subject_group_ids, object_group_ids, confidence and segments. Participants retain exact text, full-word character span and saved group_ids. Null subjects or empty links are retained if they are present in the current saved annotation; no inferred participants are silently added.

Temporal segments exactly match the review application's displayed 5 FPS timeline. Explicit human intervals take precedence; otherwise the current webpage's model-to-stored-frame display conversion is applied. Every segment contains only original source-filename frame numbers with both endpoints inclusive:

{"start_frame": 738, "end_frame": 771}

These are original filename numbers, not sampled-frame list positions. Multiple intervals are retained separately. Each endpoint belongs to frame_files_json. Seconds, review state and temporal-source fields are not included in exported action segments.

Data sources

This is a derived annotation release based on VIPOSeg media and segmentation annotations. Captions, text-to-object links, actions and temporal annotations extend the original dataset. Consult the original dataset's access, citation and licensing terms when obtaining or using its media and annotations.