Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

YAML Metadata Warning:The task_categories "anomaly-detection" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other

seg-gap-iad-data

Pre-computed anomaly energy maps for reproducing all experiments in:

Hu Xiaojun. "The Segmentation Gap in Industrial Anomaly Detection: A Halo-Plateau Diagnosis." Pattern Recognition Letters (under review).

Code: https://github.com/knifecaojia/seg-gap-iad

Contents

This dataset stores the intermediate outputs of running three anomaly-detection method families on five IAD benchmarks (225 categories total, ~310k images). It is not the raw images — those must be obtained from the original dataset publishers (see the code repo's DATA.md).

Component Size Files
PatchCore energy maps (.npy, 224×224 float32) ~58 GB 305,803
Dinomaly maps (MVTec AD + VisA) ~0.8 GB
EfficientAD maps (MVTec AD, anomalib + direct) ~0.7 GB
Per-method _meta.json + _eval_summary_*.json small
dinomaly_mvtec_model.pth (trained weights) 0.25 GB 1

Total: 61.79 GB / 306,136 files.

Why pre-computed?

Re-running PatchCore + Dinomaly + EfficientAD on all 225 categories takes ~1 day of GPU time. These energy maps let you skip that step and go straight to evaluation / halo analysis / figure generation in the code repo.

Download & extract

# Download (set HF_ENDPOINT=https://hf-mirror.com if you are in mainland China)
hf download knifeandcj/seg-gap-iad-data --repo-type dataset --local-dir seg-gap-iad-data

# Extract every part into the same target dir; they reconstruct outputs/ with
# no filename collisions. Each part-NNNNN.zip is a standalone standard zip.
cd seg-gap-iad-data
mkdir ../outputs
for p in part-*.zip; do unzip -q "$p" -d ../outputs; done

The resulting outputs/ tree matches the layout the code repo's src/paths.py expects by default (or set SGAP_OUTPUTS to point at it).

Integrity

manifest.json maps every source file to its part, size, and md5. Verify after extraction:

import json, hashlib, os
m = json.load(open("manifest.json"))
# compare md5 of each extracted file against manifest

README.json has packing metadata (part count, totals, timestamp).

License

MIT. The energy maps are derived from the following datasets, each under its own license — obtain and respect those terms for any redistribution of raw images:

  • MVTec AD / MVTec AD 2 (CC BY-NC-SA / research use)
  • VisA (Apache-2.0)
  • Real-IAD / Real-IAD Variety (research use, see real-iad.org)

Citation

@article{hu2026segmentation,
  title={The Segmentation Gap in Industrial Anomaly Detection: A Halo-Plateau Diagnosis},
  author={Hu, Xiaojun},
  journal={Pattern Recognition Letters},
  year={2026},
  note={Under review}
}
Downloads last month
68