Skip to main content

Single Scene Video Anomaly Detection Metrics

Project description

Single-Scene Video Anomaly Detection (SSVAD) Metrics

This project contains evaluation protocol (metrics) for benchmarking single-scene video anomaly detection (SSVAD).

Evaluation Protocol

This is an unofficial implementation of Sec. 2.2 of A Survey of Single-Scene Video Anomaly Detection. This metric is intended only for single-scene video anomaly detection methods and is untested for other purposes.

There are 3 kind of supported outputs from SSVAD methods/ground-truths:

  • Pixel-level anomaly scores maps.
  • Bounding-boxes based outputs, it will be implicitly converted into pixel-level anomaly scores maps. Overlapping bounding-boxes will be averaged by its scores.
  • Frame-level anomaly scores.

The metrics can be categorized into 4:

  1. Track-based metric: requires pixel-level anomaly scores maps predictions and ground-truths, plus anomaly track ID for each frame in the ground-truths that contains anomalous regions (predictions does not require anomaly track ID and is ignored in the process).
  2. Region-based metric: requires pixel-level anomaly scores maps predictions and ground-truths.
  3. Pixel-level traditional metric: requires pixel-level anomaly scores maps predictions and ground-truths.
  4. Frame-level traditional metric: only require frame-level anomaly scores; does not require pixel-level anomaly scores maps predictions and ground-truths.

Each prediction output and ground-truth annotation must be a JSON file that follows data structure defined in ssvad_metrics.data_schema.VADAnnotation. Pixel-level anomaly scores maps arrays must be provided using .tiff or .npy format, containing single-precision (32-bit) floating point values ranging from 0.0 to 1.0. For ground-truths, it can be same as pixel-level predictions, or using boolean values instead. But only .npy files supported if using boolean values (as .tiff requires 32-bit floating point values).


This metrics is available via PyPI.

pip install py-ssvad-metrics


  1. Prepare ground-truth JSON files and prediction JSON files (also the pixel anomaly score map linked files for pixel-level predictions and groundtruths).

    • ssvad_metrics.data_schema.VADAnnotation can be used for the data structure reference and validator of the JSON file.
    • JSON file examples are in the samples folder.
    • For UCSD Pedestrian 1 and 2 datasets, CUHK Avenue dataset, and Street Scene dataset, we provided scripts for converting ground-truth annotation files from Street Scene dataset provided by the paper (txt files, each row contains: <filename> <track_id> <x_center> <y_center> <width> <height>). Download link is provided in the paper
  2. Example usage for single groundtruth and prediction file pair:

    import ssvad_metrics
    result = ssvad_metrics.evaluate(
  3. Example usage for multiple groundtruth and prediction file pairs:

    import ssvad_metrics
    result = ssvad_metrics.accumulated_evaluate(
  4. For more examplles, see samples folder.

Visual Inspection

We also provide tools for visual inspection for checking the quality of false positives. After installing py-ssvad-metrics, the visualizer can be used by executing ssvad-visualize or ssvad-visualize-dir. See ssvad-visualize --help or ssvad-visualize-dir --help for details and usage. Also, see ssvad_metrics.visualize or ssvad_metrics.visualize_dir for the Python API details and usage. Requires FFMPEG installation on the system and ffmpeg-python package (NOT python-ffmpeg). FFMPEG is used instead of OpenCV VideoWriter, since the OpenCV packages that are distributed in the PyPI usually does not embed FFMPEG that uis compiled with H264 codec.


  1. B. Ramachandra, M. Jones and R. R. Vatsavai, "A Survey of Single-Scene Video Anomaly Detection," in IEEE Transactions on Pattern Analysis and Machine Intelligence, doi: 10.1109/TPAMI.2020.3040591.


GPL-3.0 License. Brought to open-source by PT Qlue Performa Indonesia.


Feel free to contribute for improvements.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for py-ssvad-metrics, version 0.4.4
Filename, size File type Python version Upload date Hashes
Filename, size py_ssvad_metrics-0.4.4-py3-none-any.whl (31.3 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size py-ssvad_metrics-0.4.4.tar.gz (17.2 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page