CMD-AD is released exclusively for use in the SLoMO Competition (ECCV 2026 Workshop).
Access is granted manually after the organizers review your request. Please provide
accurate information and read the access agreement carefully before submitting.

Access Agreement

By requesting access to the CMD-AD dataset, you acknowledge and agree to the following terms:

Sole purpose. The dataset will be used exclusively for participation in the
SLoMO Competition (ECCV 2026 Workshop). Any use outside the scope of the
Competition — including but not limited to commercial use, training models for
unrelated tasks, benchmarking on other challenges, or producing derivative
datasets — is strictly prohibited.
No redistribution. I will not redistribute, republish, mirror, share, or
otherwise make the dataset (in whole or in part, in original or modified form)
available to any third party, whether inside or outside my institution.
Confidentiality. I will take reasonable measures to prevent unauthorized
access to the dataset within my organization and will restrict access to
collaborators who are themselves registered Competition participants and bound
by these terms.
Accountability. I understand that the information I provide will be reviewed
by the organizers, and that access may be revoked at any time in the event of a
violation of these terms.

I confirm that the information I am providing is accurate and that I agree to all
of the above conditions.

CMD-AD (SLoMO Competition Release)

CMD-AD is the train + validation + private test release of the Condensed Movie Dataset with Audio Descriptions, distributed for the SLoMO Competition at the ECCV 2026 Workshop.

cmdad/
├── gt_annotations/
│   ├── cmdad_ad_train.csv                # Train audio-description annotations (with text)
│   ├── cmdad_charbank_train.json         # Character bank for train movies
│   ├── cmdad_ad_val.csv                  # Validation AD timestamps only (text WITHHELD, see below)
│   ├── cmdad_charbank_val.json           # Character bank for validation movies
│   ├── cmdad_subtitle_val.csv            # Validation subtitles
│   ├── cmdad_ad_private_test.csv         # Private test AD timestamps only (text WITHHELD, see below)
│   ├── cmdad_charbank_private_test.json  # Character bank for private test movies
│   └── cmdad_subtitle_private_test.csv   # Private test subtitles
├── video_val.zip                         # 591 validation video clips (.mkv), organized by year
├── video_private_test.zip                # 430 private test video clips (.mp4), organized by year
├── actor_profiles_val.tar                # IMDb profile images for the val charbank
├── actor_profiles_private_test.tar       # IMDb profile images for the private test charbank
├── actor_profiles_train.tar              # IMDb profile images for the train charbank
└── video_train_download/                 # Scripts and URL lists to download train videos
    ├── download_train_videos.py
    ├── download_train_videos.txt         # Usage notes
    └── yt_links_train/                   # Per-year YouTube URL lists (6,637 clips total)

After downloading, extract the archives:

unzip video_val.zip
# yields: video_val/2011/... video_val/2012/... ... video_val/2019/...

unzip video_private_test.zip
# yields: video_private_test/2021/... ... video_private_test/2026/...

tar -xf actor_profiles_val.tar           # yields: profile_image_val/nmXXXXXXX.jpg
tar -xf actor_profiles_private_test.tar  # yields: profile_image_private_test/nmXXXXXXX.jpg
tar -xf actor_profiles_train.tar         # yields: profile_image_train/nmXXXXXXX.jpg

Splits at a glance

Split	Movies	Clips	ADs	AD text	Videos
Train	1,101	6,637	76,528	released	download from YouTube
Val	98	591	7,316	withheld	`video_val.zip` (.mkv)
Private test	44	430	7,479	withheld	`video_private_test.zip` (.mp4)

The three splits are disjoint: no clip and no movie (IMDb ID) is shared between them.

Validation split

The validation split is used as the evaluation benchmark for the SLoMO Competition. To keep the benchmark honest, the audio-description text is not released: cmdad_ad_val.csv contains only the ground-truth clip boundaries, without the text column. Its columns are:

Column	Description
`ad_id`	Unique identifier for the AD interval, formatted `{cmd_filename}_{NNN}` (e.g. `2014/ewEreSjPyC4_001`). Use this key when submitting your generated ADs.
`imdbid`	IMDb movie ID (`ttXXXXXXX`); joins to the character bank in `cmdad_charbank_val.json`.
`movie_title`	Human-readable movie title.
`cmd_filename`	Condensed Movies clip path, `{year}/{video_id}`, matching the extracted `video_val/{year}/{video_id}.mkv`.
`cmd_clip_idx`	Index of the clip within the movie.
`start`, `end`	Start/end time of the AD interval, in seconds, relative to the clip.
`duration`	Interval length in seconds (`end - start`).

Participants are expected to generate the AD for each timestamped clip (keyed by ad_id); the held-out reference texts are used by the organizers to score submissions.

All other validation artefacts (videos, subtitles, character bank, profile images) are provided in full so participants can develop and evaluate their systems locally.

Private test split

The private test split is the final leaderboard benchmark for the SLoMO Competition. It covers 44 films (430 clips, 7,479 AD intervals) that appear in neither the train nor the validation split — no shared clips and no shared IMDb IDs.

As with validation, the audio-description text is withheld: cmdad_ad_private_test.csv carries the same eight columns as cmdad_ad_val.csv (ad_id, imdbid, movie_title, cmd_filename, cmd_clip_idx, start, end, duration) and no text column. Submit your generated ADs keyed by ad_id.

Everything else is provided in full:

Artefact	Contents
`video_private_test.zip`	430 clips, `video_private_test/{year}/{video_id}.mp4`, years 2021–2026
`cmdad_subtitle_private_test.csv`	13,040 subtitle lines, same columns as the val subtitles
`cmdad_charbank_private_test.json`	44 movies, 2,720 cast entries (2,672 unique actors)
`actor_profiles_private_test.tar`	859 IMDb profile images, `profile_image_private_test/nmXXXXXXX.jpg`

Train split

For licensing reasons, the train video clips themselves are not hosted here. Instead, we provide a downloader that pulls the currently valid 6,637 train clips directly from YouTube into video_train/{year}/{video_id}.{mp4,mkv}, mirroring the val layout:

cd video_train_download
python download_train_videos.py \
    --links-root yt_links_train \
    --video-root ../video_train

Requirements:

yt-dlp >= 2025 and ffmpeg on PATH
deno on PATH (or pass --deno-bin /path/to/deno) — needed for YouTube's n-signature challenge; without it ~5–15% of clips will spuriously fail

The script is resume-safe and can be re-run after a partial failure. See download_train_videos.txt for the full set of options (cookies for authenticated requests, per-year filtering, dry-run, rate-limit recovery, etc.). Expect the train videos to occupy ~35–45 GB after a full download.

Character bank format

cmdad_charbank_{train,val,private_test}.json maps each IMDb movie ID (ttXXXXXXX) to its cast list:

{
  "tt0042208": [
    {"id": "nm0001330", "name": "Sterling Hayden", "role": "Dix Handley"},
    {"id": "nm0129894", "name": "Louis Calhern",   "role": "Alonzo D. Emmerich"},
    ...
  ],
  ...
}

Character IDs use the standard IMDb nmXXXXXXX format. For each character ID, the corresponding profile photo (when available on IMDb) is provided as profile_image_{train,val,private_test}/<nm_id>.jpg after extracting the matching actor_profiles_{train,val,private_test}.tar. Not every cast member has a photo — coverage is roughly a third of the character bank — so treat a missing file as expected rather than as an error.

Intended use

This release is intended solely for the SLoMO Competition. See the access agreement above for the full terms.

Citation

Please cite both the original Condensed Movies dataset and the CMD-AD release:

@InProceedings{Bain20,
  author       = "Max Bain and Arsha Nagrani and Andrew Brown and Andrew Zisserman",
  title        = "Condensed Movies: Story Based Retrieval with Contextual Embeddings",
  booktitle    = ACCV,
  year         = "2020",
}

@InProceedings{Han24,
  author       = "Tengda Han and Max Bain and Arsha Nagrani and G{\"u}l Varol and Weidi Xie and Andrew Zisserman",
  title        = "{AutoAD III}: The Prequel -- Back to the Pixels",
  booktitle    = CVPR,
  year         = "2024",
}

Contact

For more information, please contact the SLoMO workshop organisers at eccv_slomo [at] googlegroups [dot] com.

Downloads last month: 80

Total file size:

21.2 GB

Access to CMD-AD is restricted to SLoMO Competition participants