ID-VTG: Image-Disambiguated Video Temporal Grounding

ID-VTG is a benchmark for Image-Disambiguated Video Temporal Grounding. Each query combines a text description and a reference image. The task is to localize the temporal segment in which the specific instance depicted in the reference image performs the action described by the text query.

Repository Structure

ID-VTG/
├── README.md
├── LICENSE
├── NOTICE.md
├── CITATION.cff
├── CHANGELOG.md
├── IDVTG-Gym/
│   ├── annotations/
│   │   ├── idvtg_gym_train.json
│   │   └── idvtg_gym_val.json
│   ├── images/
│   │   ├── train.zip
│   │   └── val.zip
│   ├── videos/
│   │   ├── shard_0000.tar
│   │   ├── shard_0001.tar
│   │   ├── shard_0002.tar
│   │   ├── shard_0003.tar
│   │   ├── shard_0004.tar
│   │   ├── shard_0005.tar
│   │   ├── shard_0006.tar
│   │   ├── shard_0007.tar
│   │   ├── shard_0008.tar
│   │   └── shard_0009.tar
│   └── features/
│       └── clip_vit_l14/
│           ├── video_features_2fps.zip
│           └── image_features.zip
├── IDVTG-InternVid/
│   ├── annotations/
│   │   ├── idvtg_internvid_train.json
│   │   └── idvtg_internvid_val.json
│   ├── images/
│   │   ├── idvtg_internvid_train_images/
│   │   └── idvtg_internvid_val_images/
│   ├── videos/
│   │   └── <video_id>.mp4
│   └── features/
│       └── clip_vit_l14/
│           ├── video_features_2fps.zip
│           ├── train_image_features.zip
│           └── val_image_features.zip
└── IDVTG-Web/
    ├── annotations/
    │   └── idvtg_web_test.json
    ├── images/
    │   └── images.zip
    ├── videos/
    │   └── videos.zip
    └── features/
        └── clip_vit_l14/
            └── video_features_2fps.zip
            └── image_features.zip

Precomputed CLIP Features

For convenience and reproducibility, the repository provides precomputed video and reference-image features extracted with pretrained CLIP ViT-L/14 weights.

For all released video features, videos were sampled at 2 frames per second (2 FPS) before CLIP feature extraction. The resulting features are stored as NumPy (.npy) files and distributed as ZIP archives.

The released feature archives are:

IDVTG-Gym/features/clip_vit_l14/video_features_2fps.zip
IDVTG-Gym/features/clip_vit_l14/image_features.zip

IDVTG-InternVid/features/clip_vit_l14/video_features_2fps.zip
IDVTG-InternVid/features/clip_vit_l14/train_image_features.zip
IDVTG-InternVid/features/clip_vit_l14/val_image_features.zip

IDVTG-Web/features/clip_vit_l14/video_features_2fps.zip
IDVTG-Web/features/clip_vit_l14/image_features.zip

IDVTG-Gym

IDVTG-Gym focuses on fine-grained and compositionally ordered gymnastics actions involving athletes with similar visual appearances.

Annotations:

IDVTG-Gym/annotations/idvtg_gym_train.json
IDVTG-Gym/annotations/idvtg_gym_val.json

Reference images:

IDVTG-Gym/images/train.zip
IDVTG-Gym/images/val.zip

Videos are distributed as TAR shards:

IDVTG-Gym/videos/shard_0000.tar
...
IDVTG-Gym/videos/shard_0009.tar

Precomputed CLIP ViT-L/14 features:

IDVTG-Gym/features/clip_vit_l14/video_features_2fps.zip
IDVTG-Gym/features/clip_vit_l14/image_features.zip

Inspect the member naming convention of a video shard with:

tar -tf IDVTG-Gym/videos/shard_0000.tar | head -n 30

IDVTG-InternVid

IDVTG-InternVid focuses on open-world videos containing diverse entities and strong temporal distractors.

Annotations:

IDVTG-InternVid/annotations/idvtg_internvid_train.json
IDVTG-InternVid/annotations/idvtg_internvid_val.json

Reference images:

IDVTG-InternVid/images/idvtg_internvid_train_images/
IDVTG-InternVid/images/idvtg_internvid_val_images/

Videos are distributed as individual MP4 files:

IDVTG-InternVid/videos/<video_id>.mp4

Precomputed CLIP ViT-L/14 features:

IDVTG-InternVid/features/clip_vit_l14/video_features_2fps.zip
IDVTG-InternVid/features/clip_vit_l14/train_image_features.zip
IDVTG-InternVid/features/clip_vit_l14/val_image_features.zip

Symbolic links used on the authors' servers are only a local storage mechanism. Published media and feature archives contain the actual file contents.

IDVTG-Web

IDVTG-Web is the out-of-distribution test subset constructed from web videos.

IDVTG-Web/annotations/idvtg_web_test.json
IDVTG-Web/images/images.zip
IDVTG-Web/videos/videos.zip
IDVTG-Web/features/clip_vit_l14/video_features_2fps.zip
IDVTG-Web/features/clip_vit_l14/image_features.zip

Source Data and Attribution

IDVTG-Gym is derived from FineGym resources.

IDVTG-InternVid is derived from InternVid-TG or DisTime and InternVid resources.

IDVTG-Web contains an out-of-distribution evaluation subset constructed from publicly accessible web videos.

Users must cite the ID-VTG paper and the applicable upstream datasets. Users must retain source attribution recorded in the released metadata.

License

The files distributed in this repository are offered under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License, abbreviated as CC BY-NC-SA 4.0.

This license statement applies only to rights that the relevant licensors are authorized to grant. Third-party rights may still apply to individual media files.

Intended Use

This dataset is intended for non-commercial research in video temporal grounding, multimodal video understanding, image-conditioned video retrieval, temporal localization, and instance-level video grounding.

Out-of-Scope Use

The dataset must not be used for commercial applications, biometric identification, identifying private individuals, unlawful surveillance, harassment, tracking, or uses that violate applicable licenses or law.

Citation

ID-VTG has been accepted to the 34th ACM International Conference on Multimedia (ACM MM 2026).

If you use the ID-VTG dataset, benchmark, or code in your research, please cite:

@inproceedings{zheng2026idvtg,
  author    = {Minghang Zheng and Jingli Wei and Hongyi Yang and Yang Liu},
  title     = {{ID-VTG}: Image-Disambiguated Video Temporal Grounding},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
  year      = {2026},
  month     = nov,
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  location  = {Rio de Janeiro, Brazil},
  note      = {To appear},
  eprint    = {2608.20127},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi       = {10.48550/arXiv.2608.20127},
  url       = {https://arxiv.org/abs/2608.20127}
}

Users of IDVTG-Gym should also cite the FineGym paper and dataset.

Users of IDVTG-InternVid should also cite the InternVid-TG or DisTime paper and the InternVid paper and dataset.