CPO-LED Bench v2.0

CPO-LED Bench is an image-inclusive benchmark for normal-only anomaly detection in LED-screen animations. It supports temporal-order inspection, partially observed semantic-slot consistency, natural production anomalies, and source-isolated evaluation.

The complete image archive is published as a GitHub Release asset:

Download CPO-LED-Bench-v2.0-full.zip

• Archive size: 395,563,944 bytes
• SHA-256: a462f4c144b4675aaeb7ccdc819066653fd9500615843996f941226c05a3b19d
• Data payload version: 2.0.0
• Citation metadata version: 2.0.1
• Licence: CC BY 4.0

Version 2.0.1 is a metadata revision. It corrects the dataset author order and binds a persistent dataset DOI. Images, annotations, splits, manifests, evidence, model files, and the v2.0.0 archive bytes are unchanged.

• Zenodo version DOI: 10.5281/zenodo.24031281
• Zenodo concept DOI: 10.5281/zenodo.24031280
• Zenodo record: https://zenodo.org/records/24031281

This repository contains the dataset card, image index, split definitions, model-ready manifests, configuration, frozen evaluation reports, and per-file checksums. The image files themselves are distributed through the Release asset so that the repository remains lightweight.

Dataset content

Item	Count
Dynamic difference images	50,100
Static LED crop images	20,040
Total image files	70,140
Spatial base windows	5,010
Expanded temporal windows	13,802
Natural production anomalies	222
Natural test normal windows	300

Dynamic examples contain ten consecutive difference images per base window. Static examples contain one 600 x 165 grayscale center crop for each LED subimage. The image index records role, source, sequence, window, frame, label, and relative path.

Natural and stress anomalies

The natural test subset contains production-derived device-absent and static-animation observations. These events were not injected by benchmark code. Reverse, shuffle, freeze, and frame-drop variants are stored separately as deterministic temporal stress cases and are identified by the transform field.

Splits and evaluation

Normal training, normal calibration, validation, and test use source-sequence isolation. The test split must not be used for early stopping, feature selection, conformal threshold selection, or fusion-weight selection.

• splits/normal_train.csv
• splits/normal_calibration.csv
• splits/validation.csv
• splits/test.csv
• splits/domain_test.csv
• splits/split_summary.csv

manifests/ contains the expanded CPO-STCM sample manifest and the natural test image manifest. results/ contains frozen reports used by the manuscript. checksums.sha256 records SHA-256 values for every file in the full archive except the checksum file itself.

UCSD Ped2

The archive includes derived manifests, reimplemented baseline checkpoints, and frozen evaluation reports for UCSD Ped2. Raw UCSD Ped2 images are not redistributed and remain subject to the original dataset terms. The source information is recorded in external/ucsd_ped2/summary.json.

Licence

The owner-provided LED images, annotations, split definitions, manifests, and derived evidence are released under the Creative Commons Attribution 4.0 International licence. See LICENSE. Third-party datasets and software remain subject to their respective terms.

Citation

@dataset{wang2026cpcoled,
  author    = {Wang, Wanli and Chen, Huafeng and Wei, Xian},
  title     = {{CPO-LED Bench v2.0}},
  year      = {2026},
  version   = {2.0.1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24031281},
  url       = {https://doi.org/10.5281/zenodo.24031281}
}

Machine-readable citation metadata is provided in CITATION.cff.

Contact

Please use the repository issue tracker for dataset questions or reproducible errors. Author affiliations are listed in CITATION.cff.