README.md

---
license: cc-by-nc-4.0
task_categories:
- video-classification
- video-to-video
pretty_name: ASHEN Dataset
tags:
- video-anomaly-detection
- surveillance
- cctv
- temporal-annotation
---

# ASHEN Dataset

ASHEN is a CCTV/surveillance video anomaly dataset organized for weakly supervised video anomaly detection and frame-level evaluation.

## Dataset Layout

```text
anomaly/
  Arson/
  Assault/
  Burglary/
  Explosion/
  Fighting/
  Robbery/
  Shooting/
  Vandalism/
normal/
splits_v2_native/
testvidtai/
```

## Current Version

- Anomaly classes: 8
- Published native protocol: 1,600 videos
- Anomaly videos: 800 (100 per anomaly class)
- Normal videos used by the native protocol: 800
- Native training split: 1,300 videos (650 anomaly + 650 normal)
- Native test split: 300 videos (150 anomaly + 150 normal)

Only videos listed under `splits_v2_native/` belong to the official evaluation
protocol. The two auxiliary `testvidtai` videos are not part of training or
evaluation.

## Anomaly Classes

- Arson
- Assault
- Burglary
- Explosion
- Fighting
- Robbery
- Shooting
- Vandalism

## Splits and Temporal Annotations

The current protocol is stored in:

```text
splits_v2_native/
```

Important files:

- `train_anomaly.txt`
- `train_normal.txt`
- `test_anomaly.txt`
- `test_normal.txt`
- `Temporal_Anomaly_Annotation_for_Testing_Videos.txt`
- `annotation_todo.csv`
- `Temporal_Anomaly_Annotation_for_Testing_Videos.todo.txt`
- `protocol.json`

Frame-level temporal annotation is complete for all 150 anomalous test videos.
`annotation_todo.csv` and the `.todo.txt` file are retained as annotation audit
artifacts; despite their historical names, they do not contain unfinished
records. The canonical evaluation file is
`Temporal_Anomaly_Annotation_for_Testing_Videos.txt`.

Temporal intervals use frame indices with the convention:

```text
video_name label start1 end1 start2 end2
```

`-1 -1` denotes that the second interval is absent.

The temporal annotation file has exactly one row for each anomalous test video.
Normal test videos are interpreted as entirely normal and therefore do not
need rows in the temporal annotation file.

## Version Note

This release provides the cleaned, UCF-independent native split and manually
reviewed temporal annotations. Results should be reported on the fixed 300-video
test split, with model/checkpoint selection performed without using test labels.