ViCaS-Actions

Action annotations for 11,057 training videos of ViCaS. Each video has one caption. For every action in the caption, this dataset gives:

• the words in the caption that state it (character spans),
• who does it and what it is done to, linked to the caption's entities, which come with per-frame masks,
• when it happens in the video (time segments, also as frame ranges).

Not reviewed by people. The action spans were marked by GPT-5.6-sol and the time segments by Gemini 3.7 Flash. Expect errors; see Known issues.

Videos	11,057
Entity groups (each with per-frame masks)	38,464
Actions	43,454
Videos whose caption has no action	44
Actions without a time segment	517
Frame rate of masks and frame_spans	30 fps
Videos included	No, see Videos

Columns

One row per video.

column	type	meaning
video_id	string	ViCaS video id, zero-padded to 6 digits
video_name	string	video file name
height, width	int32	frame size in pixels
num_frames	int32	frame count; masks and frame_spans index frames 0 … num_frames-1
stored_fps	float32	30.0 in every row
caption	string	the caption; every character span below is an offset into it
groups_json	string (JSON)	entities with masks, see groups_json
num_groups	int32	length of groups_json
actions_json	string (JSON)	actions, see actions_json
num_actions	int32	length of actions_json
span_model, span_prompt_sha	string	model and prompt version that marked the action spans
temporal_model, temporal_prompt_sha	string	model and prompt version that gave the time segments
actions_schema	string	vicas_actions_parquet/v1

groups_json

A list. Entry i is one tracked object, and groups fields in actions_json refer to it by i.

{"char_spans": [[0, 12], [39, 41]],
 "text": ["A competitor", "He"],
 "segmentations": [{"size": [480, 854], "counts": "Yed56d>9I8J4L3N1O2N2O1..."}, null, ...]}

• char_spans are [start, end) offsets of every mention of the object in caption; text holds the mention strings.
• segmentations has one entry per frame (num_frames long). Each entry is a COCO RLE mask (compressed counts string), or null when the object is not visible in that frame.
• One mention can cover several groups: "four pillows" is a mention of four separate pillow groups. Two groups can also have the same mention text at different offsets (two "A woman"); only the offsets tell them apart.

actions_json

A list with one entry per action:

{"action_id": 1,
 "char_spans": [[42, 47]], "text": ["holds"],
 "type": "simple",
 "subject": {"char_span": [39, 41], "text": "He", "groups": [0]},
 "objects": [{"char_span": [48, 57], "text": "ski poles", "groups": [1, 2]}],
 "frame_spans": [[0, 233]], "segments_s": [[0.0, 7.8]],
 "temporal_confidence": "high"}

• char_spans / text: every contiguous piece of caption text that states the action, in caption order. A piece can be a repeated mention of the same event ("is pushing a cart … pushes it") or the second half of a split phrasal verb ("picks the bag up" → picks, up). The two cases are not marked apart.
• type: the span model's label for the wording: simple, posture, phrasal, prepositional, passive, nonvisual_object, intention, light_verb or other. It describes the caption's construction (e.g. light_verb for "performs jumps"). The span itself follows the rules below either way.
• subject: the mention of whoever does the action. groups lists the indexes into groups_json; it is empty when the mention is not an annotated entity (e.g. "the camera"). subject is null when the caption names no one (120 actions).
• objects: mentions of what the action is done to, in the same format. The entity a preposition points to counts ("is sitting on a chair" → a chair); scene regions such as "in the park" do not.
• segments_s: time segments [start, end] in seconds, as returned by the temporal model.
• frame_spans: the same segments in frames, 0-based and inclusive: [floor(start × 30), ceil(end × 30) − 1], clipped to the video. Empty when the model did not place the action.
• temporal_confidence: the temporal model's own rating: high, medium or low.

Span rules

The span model was told to mark (abridged):

• the verb group, with its auxiliaries ("is walking", "was kicked") and a genuine phrasal particle ("picks up", "turns on");
• not modals ("can swim" → swim), directions or prepositions ("falls backwards" → falls, "looks at a dog" → looks), adverbs, or nouns, including activity nouns ("playing table tennis" → playing, with no object);
• the real action instead of a verb that only frames it ("tries to jump" → jump, "keeps chasing" → chasing);
• nothing for static descriptions ("is next to the car", "is visible", "a fallen ball").

One rule was added after the run and applied to every row: "lose/keep … balance/control" is one action, and a possessive stays out ("loses his balance" → loses + balance).

How it was made

1. Captions and masks. Taken from vicas_train_bcc_rewrite_clean.json in jieyuz2/bcc, which holds rewritten ViCaS training captions with entity groups and per-frame masks. Groups with no mask in any frame were removed, so group indexes here can differ from that file. Spans were clipped to the caption.
2. Action spans. GPT-5.6-sol (reasoning effort medium) read each caption with its list of entities and returned the actions, with subject and objects as quoted caption text. Linking those mentions to groups was done by exact offset matching in code, not by the model.
3. Time segments. Gemini 3.7 Flash watched each video with the caption and the action list and returned time segments for every action. Answers outside the video were clipped to it.
4. Not included: the 1,154 training videos whose action spans were checked by hand, and 4 videos (004238, 015520, 027359, 032695) that Gemini refused to process.

Known issues

• Nothing here was reviewed by people. Span and time errors are expected.
• Light verbs. Because nouns stay out of spans, "does a backflip" is marked as does and "is playing golf" as is playing; the span alone does not say which activity it is.
• Span slips. Some passives miss their auxiliary (hit instead of is hit), and a few spans keep an adverb inside ("is suddenly hit").
• Time segments are model output. Boundaries can be off, and 517 actions have no segment.
• Captions are rewrites and occasionally get a detail wrong (e.g. a clothing colour) while the mask is on the right object. In one group mention the stored text differs from the caption slice at its char_spans in capitalisation only; the offsets are the reference.

Videos

The videos are not included. Get them from ViCaS (GitHub, Hugging Face). Masks and frame_spans assume the video decoded at 30 fps, with frame 0 first.

Loading

The full dataset is about 2.9 GB, almost all of it masks.

import json
from datasets import load_dataset
from pycocotools import mask as mask_utils

ds = load_dataset("UWGZQ/ViCaS-Actions", split="train", streaming=True)
row = next(iter(ds))
groups = json.loads(row["groups_json"])
actions = json.loads(row["actions_json"])

for a in actions:
    subject = a["subject"]["text"] if a["subject"] else None
    print(a["text"], "| subject:", subject, "| frames:", a["frame_spans"])

rle = groups[0]["segmentations"][0]        # frame 0 of the first entity, or None
if rle is not None:
    mask = mask_utils.decode(rle)          # height x width uint8

License

• Action annotations (actions_json, num_actions, span_model, span_prompt_sha, temporal_model, temporal_prompt_sha, actions_schema): CC BY 4.0.
• Captions and masks (caption, groups_json, num_groups): copied from jieyuz2/bcc, which states no license. That data is derived from ViCaS, whose Hugging Face card lists license: cc without a version. These columns stay under their sources' terms.
• Videos are not distributed here.

Citation

If you use this data, please cite ViCaS: ViCaS: A Dataset for Combining Holistic and Pixel-level Video Understanding using Captions with Grounded Segmentation.