mtastan/ASHEN-Dataset

ASHEN Dataset

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

Dataset Layout

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:

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:

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.