Overview
ICONIC-444 is a large-scale industrial image dataset captured using a prototype free-fall sorting machine. It contains more than 3.1 million RGB images spanning 444 fine-grained classes across food and non-food categories. The dataset is designed specifically for research on out-of-distribution (OOD) detection, robustness, and fine-grained visual recognition in industrial environments.
Dataset Availability
ICONIC-444 is publicly available for non-commercial research use. The dataset is distributed as per-class ZIP archives and accessed through a lightweight, versioned metadata interface.
The entry point to the dataset is the file
meta.json
,
which specifies the latest available dataset version as well as version-specific
metadata.
For each dataset version (e.g. v1.0.0), the metadata specifies a
base_url and the name of a blobs directory, as well as
the filenames of the corresponding manifest, README, and license files.
The full URL of the manifest is obtained by concatenating the base_url
with the manifest filename specified in meta.json.
The manifest enumerates all available classes and provides integrity information
(file sizes and SHA256 checksums) for each per-class archive.
For each class, the download URL of the corresponding archive is constructed as
base_url / blobs / <category>_<group>_<class>_<sha256>.zip,
where the category, group, class name, and SHA256 checksum are defined in the
manifest.
Individual class archives can therefore be downloaded directly using the URLs
provided in the manifest.
For reproducible and reliable usage, we strongly recommend using the official download script provided in the ICONIC-444 GitHub repository , which automatically resolves the metadata, verifies file integrity, and extracts the data into the original category / group / class hierarchy.
Summary:
meta.json → base_url + manifest → base_url / blobs / per-class ZIP archives
Dataset documentation and license:
README.txt
·
LICENSE.txt
License Terms and Attribution
The ICONIC-444 dataset is released under the Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International License (CC BY-NC-SA 4.0).
Use of the dataset for commercial purposes is not permitted under this license. If you are uncertain whether your intended use qualifies as non-commercial, or if you require a commercial license, please contact us.
For questions, licensing inquiries, or further information, please contact us at office@kestrel-eye.com.
If you use ICONIC-444 in academic or scientific work, please cite our paper:
@inproceedings{krumpl2026iconic444,
title = {ICONIC-444: A 3.1-Million-Image Dataset for OOD Detection Research},
author = {Krumpl, Gerhard and Avenhaus, Henning and Possegger, Horst},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2026}
}