AnyGroundBench Dataset

AnyGroundBench is a benchmark-style collection for temporal, spatial, and spatio-temporal grounding experiments.

This release is organized by domain. Metadata and split files are included for all domains. Video files are included only for the newly captured American Football and Mouse data. Videos from third-party datasets are not redistributed in this release.

Paper

AnyGroundBench: A Specialized-Domain Benchmark for Video Grounding in Vision-Language Models

Contents

.
|-- README.md
|-- LICENSE
`-- data/
    |-- animal/
    |-- industry/
    |-- safety/
    |-- sports/
    `-- surgery/

Each domain directory follows this layout:

data/<domain>/
|-- meta-data/
|   |-- s_train.json
|   |-- s_test.json
|   |-- t_train.json
|   |-- t_test.json
|   |-- st_train.json
|   `-- st_test.json
|-- data_splits/
|   |-- train.txt
|   `-- test.txt
`-- videos/
    |-- clips/
    `-- clips4spatial/

The safety domain also includes filename_mapping.json.

Domains

Video Layout

Temporal and spatio-temporal videos are stored under:

data/<domain>/videos/clips/<video_name>.mp4

Spatial clips are stored under:

data/<domain>/videos/clips4spatial/<video_name>.mp4

The old videos/0/ layout is no longer used.

In this release, data/animal/videos/ contains only Mouse videos, and data/sports/videos/ contains only American Football videos. The other domains include empty clips/ and clips4spatial/ directories for local reconstruction.

Redistribution Policy

This repository releases videos only for the newly captured American Football and Mouse datasets. For all third-party datasets, this repository provides metadata and split files only; users must obtain source videos separately under the original provider terms.

Users are responsible for complying with the licenses and usage terms of the original datasets.

License

The newly captured American Football videos, Mouse videos, and metadata released in this repository are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Third-party datasets remain governed by their original licenses and terms.