DENSEWORLD-115K Video Clips

115,687 video clips (4-10 s each, ~8.6 s mean) segmented from 714 Indian street videos (walking tours, driving tours, drone views) across 22 cities, for evaluating video foundation models on populous, crowded and chaotic Global South urban scenes.

Dataset Description

Property	Value
Format	WebDataset (TAR shards)
Total Clips	115,687
Shards	116 x ~1 GB TAR files
Source Videos	714
Duration Range	4.0 s - 10.0 s
Mean Duration	~8.6 s
Total Size	121.5 GB
Total Hours	~276 hours source footage
Cities	22 (6 tier-1, 15 tier-2, Goa; plus monument sites)

Format

WebDataset TAR shards with paired mp4 + json files:

data/
├── train-00000.tar
│   ├── 000000.mp4    # video clip
│   ├── 000000.json   # metadata (video_id, section, duration, etc.)
│   ├── 000001.mp4
│   ├── 000001.json
│   └── ...
├── train-00001.tar
└── ...

Loading

from datasets import load_dataset

# Streaming (recommended — no local download)
ds = load_dataset("anonymousML123/denseworld-115k", streaming=True)
for sample in ds["train"]:
    video = sample["mp4"]
    metadata = sample["json"]

Processing Pipeline

1. Download: yt-dlp at 480p (714 videos from YouTube)
2. Scene Detection: PySceneDetect ContentDetector (threshold=15.0)
3. Greedy Split: Scene-aware splitting, 4-10s clips, libx264 CRF 28

Metadata Fields

Each clip's JSON sidecar contains:

Field	Description
video_id	YouTube video ID
section	Geographic section (e.g., tier1/mumbai/drive)
tier	City tier (tier1, tier2, goa, monuments)
city	City name
tour_type	Tour type (walking, drive, drone, rain)
duration_sec	Clip duration in seconds
size_mb	File size in MB
source_file	Original clip filename

Intended Use

• Evaluate video foundation models on dense, heterogeneous urban scenes
• Test geographic transfer of video foundation models
• Research on non-Western video understanding

Citation

@dataset{denseworld115k,
  title={DENSEWORLD-115K: Indian Street Video Clips for Video Foundation Model Evaluation},
  author={Anonymous},
  year={2026},
  url={https://huggingface.co/datasets/anonymousML123/denseworld-115k}
}

License

CC-BY-4.0 (Creative Commons Attribution 4.0)

Original videos sourced from YouTube walking/driving/drone tour channels.