DENSEWORLD-115k

A large-scale video benchmark of populous, crowded, and chaotic Global South urban environments, used to study world models (JEPA) under soft spatial boundaries, extreme agent heterogeneity, persistent occlusion, and rapid social negotiation.

• 115,687 clips (4–10 s each)
• 714 long-form source videos across 22 Indian cities
• Drive-through, walk-through, and aerial (drone) viewpoints; markets, ghats, junctions, flyovers, beaches, and more

⚠️ This is a metadata-only dataset (no videos included)

The source clips are derived from public YouTube videos and remain under their creators' copyright. We therefore do not redistribute any video files. Instead, this repository ships the source video list, a clip manifest, and a deterministic reconstruction script so you can rebuild the exact clips locally from YouTube. This mirrors the standard practice of YouTube-derived datasets (e.g. Panda-70M, HD-VILA-100M, HowTo100M).

Repository contents

denseworld-115k/
├── README.md                     # this card
├── requirements.txt              # yt-dlp, scenedetect[opencv]
├── reconstruct.py                # download → scene-detect → split → rebuild the 115,687 clips
├── sources.json                  # 714 YouTube source videos (id, url, section, n_clips, title)
├── clips.csv                     # 115,687-clip manifest (clip_key, section, base_video_id, chunk, duration…)
└── data/
    └── data_prep/
        ├── clip_durations.json   # per-video clip manifest (ground truth for verification)
        ├── city_matrix.json      # per-city × capture-type coverage counts
        └── word_frequency.json   # source-title word frequencies + scene taxonomy mapping

Reconstructing the clips

pip install -r requirements.txt          # yt-dlp + PySceneDetect
# install ffmpeg + ffprobe (brew install ffmpeg  /  apt-get install ffmpeg)

python reconstruct.py --out ./denseworld_clips --limit 1   # smoke test (1 video)
python reconstruct.py --out ./denseworld_clips             # full rebuild (all 714 videos)
python reconstruct.py --out ./denseworld_clips --verify-only   # check counts vs the manifest

The pipeline is deterministic and reproduces the released clips exactly:

1. Download each source video at 480p (yt-dlp).
2. Detect scene boundaries (PySceneDetect ContentDetector, threshold 15.0).
3. Greedy-split into contiguous 4–10 s clips at scene boundaries.
4. Encode each clip (ffmpeg libx264 CRF 28, AAC 128k).

Clips are written to <out>/<section>/<video_id>-<NNN>.mp4 (e.g. goa/walking/04YKvC8kAgI-000.mp4), matching the clip_keys in clips.csv.

Some source videos may become unavailable over time (deleted or made private on YouTube); reconstruct.py skips these and reports the shortfall in its verification summary. Three very long videos were originally cut in fixed windows rather than whole-video scene detection, so their clip boundaries may differ slightly (<1% of the dataset).

Data fields

sources.json → videos: [ { … } ]

field	description
id	11-char YouTube video ID (source URL = https://www.youtube.com/watch?v=<id>)
url	full YouTube watch URL
sections	list of city/capture-type sections this video contributes to (e.g. goa/walking)
n_clips	number of clips produced from this video
n_chunks	>0 only for the 3 pre-chunked long videos
title, category	source title and collection category (drive_tours, walking_tours, drone_views, tier2_cities)

clips.csv (115,687 rows)

column	description
clip_key	section/video_id/file — canonical clip identifier
section	city/capture-type (e.g. mumbai/drive)
base_video_id	the 11-char YouTube ID
chunk	fixed-window chunk index for the 3 long videos, else empty
video_id	manifest key (base_video_id or base_video_id-<chunk>)
clip_index	clip order within its video/chunk
duration_sec, size_mb, status	encoded clip stats

License & responsible use

• The metadata and code in this repository are released under CC-BY-4.0.
• The videos are NOT included and are not covered by this license — they remain the property of their original YouTube uploaders and are subject to YouTube's Terms of Service. Use the reconstructed clips for research purposes and in accordance with those terms.
• Takedown: if you are a rights holder and want a source removed from sources.json, please open an issue on this repository and we will remove it.

Citation

@article{wanaskar2026factorjepa,
  title  = {FactorJEPA: Factorizing Monolithic Futures into Layout--Agent--Interaction
            Channels for Crowded and Chaotic Global South Urban Worlds},
  author = {Wanaskar, Kapil and Jena, Gaytri and Chadha, Aman and Jain, Vinija
            and Sharma, Vasu and Das, Amitava},
  year   = {2026},
  note   = {Preprint. Update with arXiv ID when available.}
}