Real-IAD-MVN

This release contains 20 category directories, 8,700 sample directories (ng: 3700, ok: 5000), and 104,339 image files. Counts refer to the files indexed in this release.

Release organization

Real-IAD-MVN/
├── README.md
├── ReadMe.txt
├── LICENSE.txt
├── .gitattributes
├── dataset_info.json
├── dataset.json
├── realiad_mvn_jsons.zip
├── metadata/
│   └── samples.jsonl
├── format/
├── paper_comparison.json
├── quality_report.json
├── packaging_report.json
├── SHA256SUMS
├── RELEASE_READY.json
├── verify_release.py
└── realiad_mvn_raw/
    ├── dc_power_connector.zip
    ├── fork_crimp_terminal.zip
    └── ... (one ZIP per category)

The archive-based release organization follows the Real-IAD D³ repository.

Paper and dataset scope

The associated paper is Real-IAD MVN: A Multi-View Normal Vector Dataset and Benchmark for High-Fidelity Industrial Anomaly Detection. Section 3.3 describes 20 object categories, 8,450 samples (5,000 normal and 3,450 abnormal), five RGB images and five surface-normal maps per sample, with pixel-level masks where available for abnormal samples.

This release preserves all files scanned from the supplied source directories. Its counts are shown above and in dataset_info.json. paper_comparison.json compares the source directory counts against the paper's sample counts. Category-specific defect directory names are retained verbatim; their count is reported as a directory count.

Data partitioning

This release does not introduce a train/validation/test partition or a split field. The cited paper version does not specify the partition ratios, sample assignments or random seed needed to reproduce an official partition. Each sample directory and all its views remain together in one record.

Extraction and file paths

Extract each category ZIP into realiad_mvn_raw/. Each ZIP contains a directory named after its English category identifier. Within that directory, the original status, defect, sample, view directories and image filenames are retained. Images are copied without resizing, conversion or recompression.

Extract realiad_mvn_jsons.zip into the dataset root. It contains realiad_mvn_jsons/<category>.json, one index per category. path is relative to the extracted realiad_mvn_raw/ directory. source_path records the path in the original dataset. Both use / separators.

dataset.json is the complete aggregate JSON, with dataset statistics in dataset_info and all sample records in samples. metadata/samples.jsonl contains one sample record per line. Category records in the aggregate JSON/JSONL also carry category. Images outside sample directories are retained in a separate files_outside_sample_directories array in dataset.json and in the category index.

JSON index fields

These are the index fields used by this release:

Field	Meaning
category	English category identifier used in this release
source_directory_name	Original category directory name
samples	Records grouped by original S<number> directory
sample_id	Unique identifier based on category, status, defect and sample directory
sample_directory	Sample directory relative to realiad_mvn_raw/
sample_directory_name	Original sample directory basename
status_directory_name	Original ok or ng directory name, when present
defect_directory_name	Original intervening defect directory, when present
views	Files grouped by observed C<number> directory
view_directory_name	Original view directory name
files	All image files in that group, including mask-named images
path	Released file path relative to realiad_mvn_raw/
source_path	Corresponding original relative file path
bytes	Byte length read when the image was packaged
sha256	SHA-256 of the exact original image bytes
filename_group	Literal filename match, included only when a rule matches

Images outside a view or sample directory are retained in files_outside_view_directories or files_outside_sample_directories, respectively, when present. Directory identifiers are retained as recorded. Filename rules match mask, RGBL05/a separated RGB token, or a separated PS token; mask matching takes precedence. These matches are file-discovery aids. Every image is retained, including files without a match. Multiple matches are kept as separate file entries.

import json
from pathlib import Path

root = Path("Real-IAD-MVN")
index = json.loads((root / "realiad_mvn_jsons/dc_power_connector.json").read_text(encoding="utf-8"))
sample = index["samples"][0]
file_path = root / "realiad_mvn_raw" / sample["views"][0]["files"][0]["path"]
print(file_path)

Measured image properties

format/format_measurements.json summarizes measured headers and all decoded pixels of mask-named files. format/image_headers.jsonl records width, height, file format, Pillow mode/bands, and stored PNG/JPEG bit depth when read from the file. format/mask_pixel_values.jsonl records exact decoded mask pixel values; large value sets are saved as .npy files under format/pixel_values/.

Dimension comparisons report equality of image dimensions. They do not establish geometric alignment. No normal-vector decoding formula, camera orientation, coordinate axes, mask threshold or pixel-label interpretation is assigned by this packaging process.

quality_report.json records filename matches, view directory counts, duplicate filenames by group, dimension differences and file read failures. All source images are preserved. paper_comparison.json records factual count comparisons; image counts and sample-directory counts are distinct units.

Integrity verification

SHA256SUMS contains SHA-256 checksums for the published files other than the checksum list itself. Each image also has its own digest in the JSON indexes. RELEASE_READY.json is written only after archives and JSON references pass structural verification.

After downloading or copying the release, run:

python3 verify_release.py --root /path/to/Real-IAD-MVN-hf

This checks every published-file checksum, category archive entries, their byte lengths, sample IDs, JSON coverage and the absence of introduced data partitions. No original server files are required.

Categories

Category	Sample directories	Image files
dc_power_connector	450	5498
fork_crimp_terminal	400	4657
ethernet_connector	400	4748
fuse_holder	400	4742
common_mode_filter	450	5477
crimp_st_cable_mount_box	450	5410
connector_housing_female	450	5487
lego_pin_connector_plate	450	5331
knob_cap	400	4643
lattice_block_plug	450	5164
humidity_sensor	450	5493
ferrite_bead	350	3999
power_jack	450	5348
telephone_spring_switch	500	6250
miniature_lifting_motor	350	3982
purple_clay_pot	450	5500
headphone_jack_socket	500	6243
lego_propeller	400	4736
limit_switch	450	5456
audio_jack_socket	500	6175

License

The released dataset is licensed under CC BY-NC-SA 4.0. See LICENSE.txt for the full license text.

Upload

See UPLOAD.md for uploading the complete generated release directory to a Hugging Face dataset repository.