VQS-4K

Find an object in a video, given one picture of it from somewhere else — and segment it everywhere it appears.

You get an untrimmed video and a visual query: a single image plus a mask of the target, taken from outside that video. The task is to output a pixel mask for the object in every frame where it appears — and nothing where it doesn't.

This differs from the two neighbouring tasks:

• vs. VQL (visual query localization): VQL returns one bounding box on the target's last appearance. VQS returns pixel masks for all appearances.
• vs. VOS (video object segmentation): VOS takes its reference mask from the video's own first frame. Here the query is external, so lighting, pose, scale and background can differ sharply from anything in the video.

Dataset for the NeurIPS 2026 Datasets & Benchmarks paper Towards Visual Query Segmentation in the Wild. Code and the VQ-SAM baseline: https://github.com/VAIL-Fan/VQS.

Video–query pairs	4,111
Search-video frames	1,304,473 @ 6 fps
Annotated masks	222,173 + 4,111 query masks
Categories	222 fine / 19 coarse
Size	76 GB

Quick start

Try it without downloading anything — 222 samples, one per category, render right in the viewer:

from datasets import load_dataset
ds = load_dataset("bf0820/VQS-4K", "preview", split="train")
ds[0]["query_overlay"]   # query image with the target highlighted

Grab one real sequence (~19 MB):

from huggingface_hub import hf_hub_download
import tarfile

seq = "mammal-mouse-set_8"
for kind in ("images", "masks"):
    p = hf_hub_download("bf0820/VQS-4K", f"{kind}/{seq}.tar", repo_type="dataset")
    tarfile.open(p).extractall(f"VQS-4K/{kind}")
# -> VQS-4K/images/mammal-mouse-set_8/00000.jpg ...

Grab a whole split with the helper script:

python scripts/download.py --split test --protocol tiou --out VQS-4K --unpack

How the data is laid out

Frames are packed one tar per sequence, so you can take a single video instead of all 76 GB:

images/<seq_id>.tar    4,111 archives -> <seq_id>/00000.jpg, 00001.jpg, ...
masks/<seq_id>.tar     4,111 archives -> <seq_id>/00000.png, ...
splits/                official split files, one sequence id per line
metadata.csv           per-sequence index (category, splits, frame counts, annotated range)
dataset_stats.json     totals + the full 222-category list
scripts/               download helper and a ready-made PyTorch Dataset
preview/               the 222-sample preview config

Every tar extracts into its own <seq_id>/ folder, so unpacking them all into one place rebuilds the original tree.

Three things to know about the frames

1. Frame 00000 is the query, not part of the video. images/<id>/00000.jpg + masks/<id>/00000.png define the target. The search video starts at 00001.
2. Masks are sparse on purpose. A mask file exists only for frames where the object is present. No mask file means the object is absent, and the correct prediction there is an empty mask — this is the whole point of the untrimmed setting, so don't treat missing masks as unlabeled.
3. Masks are binary PNGs (non-zero = target); frames are JPEG.

Sequence IDs

Format is <supercategory>-<category>-set_<n>, e.g. mammal-mouse-set_8. Names are plain snake_case ASCII — no spaces, so paths are safe to use unquoted.

Careful: category alone is not unique — mouse exists under both electronic and mammal. Use category_id (<supercategory>-<category>) from metadata.csv as the key. That is what gives 222.

Splits

Two protocols, each covering all 4,111 sequences. They answer different questions — report which one you used.

1. Standard split — use this one by default

splits/train_t-IoU.txt (3,289) / splits/test_t-IoU.txt (822) · column split_tiou

This is the split the paper evaluates on, and the one to use for comparable numbers. Classes overlap between train and test, so it measures ordinary in-distribution performance.

Coverage is near-total but not exact: test covers 217 of 222 classes and train 221 of 222. The five classes with no test sequence are aquatic_animal-urchin, human-rider, human-skydiver, human-swimmer, mammal-pig; sport_equipment-ice_hockey has no train sequence. Most of these are classes with very few sequences overall (urchin, swimmer and ice_hockey have exactly one).

2. Unseen-class split — for zero-shot / generalization

splits/train_t-IoU_by_class.txt (3,301) / splits/test_t-IoU_by_class.txt (810) · column split_by_class

Split along class boundaries: 180 classes in train, 42 in test, with zero overlap. No test category is ever seen during training, so this measures generalization to novel object categories rather than in-distribution accuracy. Numbers from this split are not comparable with the standard split — they are a harder, different task.

The 42 held-out classes span 14 supercategories (aircraft, aquatic_animal, bird, clothing_accessory, electronic, food, insect, kitchenware, mammal, personal_care, reptile, sport_equipment, vehicle, watercraft).

Categories

Supercategory	Seqs		Supercategory	Seqs
human	461		vehicle	143
daily_item	344		tool	135
electronic	342		retail_item	104
kitchenware	337		reptile	100
sport_equipment	332		toy	80
clothing_accessory	324		aircraft	68
personal_care	298		watercraft	58
mammal	275		food	40
aquatic_animal	240
insect	218
bird	212

Supercategory folder prefixes are the snake_case form of the paper's class names, so the figure above and the directory layout agree (clothing_accessory = "clothing & accessory", daily_item = "daily item", retail_item = "retail item", sport_equipment = "sport equipment").

Note on sequence IDs

IDs here are cleaned up relative to the paper's supplementary and the original Kaggle release: supercategory prefixes now follow the paper's class names, three class spellings are corrected (natwhal → narwhal, saiulplane → sailplane, sparyer → sprayer), and the one class containing a space is fixed (electric_impac_ wrench → electric_impact_wrench).

The frames, masks, splits and all statistics are unchanged — only names differ. Every row of metadata.csv carries a legacy_seq_id column with the original ID, so results reported against the earlier naming can be mapped across exactly:

import pandas as pd
m = pd.read_csv("metadata.csv")
old2new = dict(zip(m.legacy_seq_id, m.seq_id))

More samples

Annotation

Five steps: category and search design → collection from YouTube by keyword → manual screening to keep only videos with a valid clip + external-query pair → expert labeling of every temporal segment containing the target plus pixel masks within them → multiple rounds of inspection and refinement. Over 8K candidate videos were filtered down to 4,111.

Evaluation

Scored on mask accuracy together with a temporal IoU (t-IoU) term, so a method has to get both where and when the target appears. Because masks are sparse, frames with no ground-truth mask count: predicting a mask on an absent object is penalized. Exact protocol and the evaluation code are at https://github.com/VAIL-Fan/VQS.

License and intended use

Source videos were collected under Creative Commons licenses and are distributed for research purposes only. This release is CC BY-NC 4.0.

Source licenses can change. Rights holders can open an issue on GitHub or a discussion here and the maintainers will act.

The videos are in-the-wild YouTube footage and some contain people (the human supercategory includes cyclists, skaters, workers, passengers). Annotations mark object instances only — no identity, demographic or other personal attributes are labeled. Do not use this data for person identification, surveillance or biometric applications.

Citation

@inproceedings{fan2026vqs,
  title     = {Towards Visual Query Segmentation in the Wild},
  author    = {Fan, Bing and Li, Minghao and Zhang, Hanzhi and Dong, Shaohua and
               Mareedu, Naga Prudhvi and Liu, Xiaoqiong and Shi, Weishi and
               Feng, Yunhe and Huang, Yan and Fan, Heng},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS),
               Datasets and Benchmarks Track},
  year      = {2026}
}

Maintained by the authors, Department of Computer Science and Engineering, University of North Texas. Questions: the Community tab here, or github.com/VAIL-Fan/VQS.