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

VIPOSeg Caption, Action and Temporal Annotations — TRAIN Final

This release contains all 2,800 annotated TRAIN videos available in the local VIPOSeg review application at export time. It preserves the latest saved caption, entity, object-grounding, action, participant-link and temporal edits, regardless of confirmation state. Review states, review notes and confirmation metadata are omitted from the annotation table. VAL is not included.

Review coverage: Before export, every sample labeled repaired and every sample in the medium confidence queue was checked. These labels describe review coverage in the source workflow; they are not stored as row-level review-state fields in the annotation table.

The eight-column schema follows VIPOSeg-caption-action-temporal-train. That repository is retained separately. This release is a saved snapshot: later changes in the review application are not automatically synchronized. “Final” is the release name; it does not imply that every annotation was manually confirmed or that automatic validation established semantic correctness.

Files and coverage

Path	Contents
train/verification.parquet	2,800 rows of captions, entities, span links, masks, actions and temporal segments
train/manifest.json	Counts, checksums, source revision and coordinate conventions
README.md	Dataset description and schema

Rows follow ascending numeric video-folder prefix, then full folder name. There are 2,801 folders in the original TRAIN media directory, but 590_6E8fnQ4y9OE has no annotation record in the current review application or the reference Parquet. It is not given an invented caption or included as an empty annotation row. The actual annotated corpus contains 2,800 videos.

Statistics

Item	Count
Video/caption rows	2,800
Stored source-frame references	66,930
Entity groups	19,381
Object tracklets	19,296
Present object masks	390,491
Unique text-span links	16,403
Actions	6,988
Temporal segments	7,538
Actions with multiple segments	410

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

Parquet schema

Every row contains these eight string columns. Nested values are serialized JSON arrays for compatibility with the reference dataset.

Column	Contents
video_id	Original full video-folder identifier
split	Always train
text	Latest saved complete caption
frame_files_json	Ordered relative paths to every stored source frame
groups_json	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 to source frames
actions_json	Action text/spans, type, mentions, participants, confidence and temporal segments

Entities and text spans

An entity group retains group_id, mark_id, category, isthing, mentions, agreement metadata and any explicit ungrounded identity. Each group additionally has tracklet_ids. Grounded objects use identifiers such as viposeg-0002; these are local to a video, not globally unique. Explicitly ungrounded groups have an empty tracklet list.

A mention retains its text, span, occurrence, confidence and collective metadata. Character spans are zero-based and half-open: [start, end). The exact substring text[start:end] must equal the mention text. The occurrence metadata is the one-based, case-sensitive occurrence in an overlapping raw substring scan; it is normalized to round-trip to the authoritative span. Word boundaries are checked separately so it embedded inside white, with, sits or guitar cannot count as a pronoun span.

Each span link contains start, end, text, group_ids and tracklet_ids. A shared span may link to several groups and objects. Overlapping person/possessive/body-part phrases remain valid when they describe distinct linked entities.

Frame paths and masks

Source-frame references use the following pattern:

JPEGImages/<video_id>/<original_filename>.jpg

RGB images are not bundled. Obtain the original TRAIN media separately and use its directory as the media root.

Each tracklet retains tracklet_id, source, source_annotation_id, category metadata, present_frames and masks. There is exactly one mask entry for every path in frame_files_json: a COCO compressed RLE object with size and string counts when the object is present, or null when absent. COCO masks use column-major encoding. Object IDs correspond to the original indexed segmentation masks at:

Annotations/<video_id>/<original_frame_number>.png

Only objects referenced by current grounded entity groups are retained. TRAIN category labels are derived from the original annotation mapping; they are not official category annotations.

Actions and temporal segments

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.

Subjects and objects retain their exact caption text, character span and group_ids. Manually saved group assignments are preserved. Individual object participants are retained even when an action has multiple objects. Component mentions preserve repeated or discontinuous action text and particles. The confidence and agreement fields describe annotation judgments and are not review or confirmation states.

Temporal segments match the review application's displayed 5 FPS timeline: explicit saved human intervals take precedence; otherwise the application's model-to-stored-frame conversion is applied. This avoids exporting different endpoints from those visible in the webpage. Every exported segment uses original image-filename frame numbers and includes both endpoints:

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

This example is from 0_wHveSGjXyDY. Frame 217 denotes 00000217.jpg, not position 217 in a sampled frame list. Frame numbers can have gaps; a segment covers the corresponding part of the ordered stored-frame list. Multiple segments are retained separately. Action intervals contain no seconds, review state or temporal-source fields.

Data sources

This is a derived annotation release based on VIPOSeg media and associated 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.