README.md

---
license: cc-by-nc-sa-4.0
language:
- en
tags:
- Video
---

# Atypical Dataset

A curated video dataset capturing atypical, anomalous, and visually unconventional human activities. Sourced from real-world, synthetic, and artistic domains, it enables research into out-of-distribution video understanding.

## Overview

The **Atypical** video dataset introduces a diverse collection of short video clips that significantly diverge from everyday behavior and appearance. Unlike standard datasets that focus on typical human actions and natural scenes, **Atypical** emphasizes the rare, the surreal, and the unexpected—drawing from both real-life occurrences and fictional or synthetic creations.

## Key Features

- **Behavioral Anomalies:** Includes clips of unintentional behavior, accidents, and social norm violations.
- **Visual Deviation:** Features surreal synthetic renderings, sci-fi environments, and staged theatrical content.
- **Genre Diversity:** Integrates real-world, cinematic, animated, and artificial domains.
- **Fine-grained Segmentation:** Clips are manually segmented into 2–10 seconds to ensure diversity and focus.
- **Multi-source Compilation:** Videos are sampled from multiple publicly available datasets and online sources, curated with consistent formatting and labeling.

## Data Composition

### Unintentional & Abnormal

- **Sources:**  
  - [Oops Dataset](https://oops.cs.columbia.edu/data/)  
  - [UCSD Ped2](http://www.svcl.ucsd.edu/projects/anomaly/dataset.htm)  
  - [CUHK Avenue](https://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html)  
  - [UCF-Crime](https://www.crcv.ucf.edu/projects/real-world-anomaly-detection-in-surveillance-videos/)

These videos showcase anomalous activities like accidents, social violations, and unexpected events.

### Surreal

- **Source:**  
  - [SURREAL Dataset](https://www.di.ens.fr/willow/research/surreal/)  
    Photorealistic synthetic human renderings performing various motions.

### Sci-fi, Animation, Theatre

- **Source:**  
  - Public YouTube videos, including:
    - Sci-fi film trailers with futuristic or supernatural elements
    - Cinematic animations with stylized visuals
    - Stage performances with exaggerated theatrical expressions

## Data Preprocessing

- **Temporal Segmentation:**  
  Long videos were manually segmented into 2–10 second clips to isolate key moments and remove redundancy.

- **Content Filtering:**  
  Non-relevant or noisy content (e.g., intros, subtitles, logos) was excluded to focus on core visual semantics.

- **Resolution Normalization:**  
  All clips were resized to a consistent format (e.g., 720p) and re-encoded using standard codecs (H.264, MP4) for compatibility.

- **Metadata Annotation:**  
  Each clip includes metadata describing the source type (e.g., `anomaly`, `surreal`, `sci-fi`), duration, and original dataset/video ID for traceability.

## Download This Repo

You can use the following code to download the entire dataset:

```python
from huggingface_hub import snapshot_download

repo_id = "july98/atypical_video_dataset"

snapshot_download(repo_id=repo_id, repo_type="dataset", token={your_token})
```

## Dataset Detail

### Project Description
Developed by: Qiyue Sun, Qiming Huang, Yang Yang, Hongjun Wang, Jianbo Jiao
License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0

### Dataset Statistics

Total Videos: 7,818 video clips spanning a wide range of atypical scenarios.
Sources:

Unintentional behaviors: 2,835 clips sourced from the Oops dataset, with an average duration of 9.77 seconds.
Abnormal scenes: 1,103 clips extracted from established anomaly detection datasets including UCSD Ped2, CUHK Avenue, and UCF-Crime, averaging 7.53 seconds per clip.
Surreal content: 1,024 clips from the SURREAL dataset, featuring synthetic human renderings with an average duration of 3.18 seconds.
Sci-fi: 1,119 clips curated from YouTube film trailers that depict futuristic or supernatural themes, averaging 4.00 seconds.
Animation: 1,058 clips taken from animated movie trailers on YouTube, averaging 4.04 seconds.
Theatre: 679 clips from stylized stage performances sourced from YouTube, with an average duration of 4.81 seconds.
Duration: The dataset comprises over 12.3 hours of video, with an overall average clip length of 5.70 seconds.

All videos have been segmented into short clips and preprocessed for consistency in resolution and format.