PolyMFO

PolyMFO is a high-resolution image dataset for foreign-object anomaly classification and binary segmentation. It contains normal samples and samples from nine foreign-object classes. Every image has a corresponding binary mask.

This Hugging Face repository contains the dataset only.
Source code for data processing, benchmark preparation, evaluation, and related research utilities is maintained at https://github.com/CaoMinhh/PolyMFO.

Dataset summary

Property	Value
Total images	575
Image resolution	2160 × 2160 pixels
Training images	460
Testing images	115
Normal images	288
Anomaly images	287
Anomaly classes	9
Split	80% training / 20% testing
Segmentation mask	Binary, 0 = background and 1 = anomaly

The dataset uses a deterministic stratified holdout split based on status and class_id, with random seed 20260904.

Dataset visualization

Representative samples from PolyMFO are shown below. Each sample includes the input image, its binary segmentation mask, and the corresponding anomaly overlay. The visualization covers the normal category and all nine foreign-object classes.

Class distribution

Label index	Class ID	Foreign-object type	Training	Testing	Total
0	—	Normal	230	58	288
1	F001	White textile thread	23	5	28
2	F002	Black textile thread	29	7	36
3	F003	Human hair	26	6	32
4	F004	Transparent plastic film fragment	15	4	19
5	F005	Cling film fragment	20	5	25
6	F006	Colored plastic film fragment	26	7	33
7	F007	Cardboard fragment	39	10	49
8	F008	Plastic flash fragment	23	6	29
9	F009	Cable tie fragment	29	7	36

Directory structure

PolyMFO/
├── .gitattributes
├── .gitignore
├── LICENSE
├── README.md
└── PolyMFO_V1/
    ├── classes.csv
    ├── final_polymfo_annotations.csv
    ├── split_info.json
    ├── visual_dataset/
    │   └── visual_dataset.jpg
    ├── training/
    │   ├── normal/
    │   ├── normal_mask/
    │   ├── anomaly/
    │   └── anomaly_mask/
    ├── testing/
    │   ├── normal/
    │   ├── normal_mask/
    │   ├── anomaly/
    │   └── anomaly_mask/
    └── npy/
        ├── x_train.npy
        ├── x_test.npy
        ├── segmentation/
        │   ├── y_train.npy
        │   └── y_test.npy
        └── classification/
            ├── y_train.npy
            └── y_test.npy

Annotation files

final_polymfo_annotations.csv

Column	Description
image_path	Image path relative to the dataset root
mask_path	Corresponding binary-mask path
split	training or testing
status	normal or anomaly
class_id	F001–F009; empty for normal images
foreign_object_type	none for normal images
mask_area_percent	Mask foreground area as a percentage of the image, rounded to four decimal places
updated_at	Annotation update timestamp, when available

Other metadata

• classes.csv maps each anomaly class ID to its foreign-object type.
• split_info.json records the split method, random seed, counts, and dataset version.

NumPy arrays

All provided arrays use float32.

File	Shape	Description
npy/x_train.npy	(460, 2160, 2160, 3)	RGB training images
npy/x_test.npy	(115, 2160, 2160, 3)	RGB testing images
npy/segmentation/y_train.npy	(460, 2160, 2160, 1)	Binary training masks
npy/segmentation/y_test.npy	(115, 2160, 2160, 1)	Binary testing masks
npy/classification/y_train.npy	(460, 10)	One-hot training labels
npy/classification/y_test.npy	(115, 10)	One-hot testing labels

The classification index order is:

0: normal
1: F001
2: F002
3: F003
4: F004
5: F005
6: F006
7: F007
8: F008
9: F009

Because the image and segmentation arrays are large, memory mapping is recommended:

import numpy as np

x_train = np.load("npy/x_train.npy", mmap_mode="r")
y_seg_train = np.load("npy/segmentation/y_train.npy", mmap_mode="r")
y_cls_train = np.load("npy/classification/y_train.npy", mmap_mode="r")

print(x_train.shape, x_train.dtype)
print(y_seg_train.shape, y_seg_train.dtype)
print(y_cls_train.shape, y_cls_train.dtype)

Code repository

This Hugging Face repository is dedicated to the PolyMFO dataset.

The source code for data processing, benchmark preparation, evaluation, and related research utilities is available at:

https://github.com/CaoMinhh/PolyMFO

For implementation details and reproducibility instructions, please refer to the GitHub repository.

Intended uses

PolyMFO is intended for non-commercial academic and scientific research, including:

• Image classification
• Binary semantic segmentation
• Anomaly detection and localization
• Industrial visual-inspection benchmarking
• Evaluation of models under high-resolution inputs
• Reproduction and comparison of research results

Commercial use is not permitted under the CC BY-NC 4.0 license.

Limitations

• The dataset contains 575 images and may not represent all real-world foreign objects, imaging conditions, or operating environments.
• Class frequencies are not identical, especially in the testing split.
• Results obtained on PolyMFO should not be interpreted as proof of generalization to other acquisition systems or industrial environments.
• Users should report the provided fixed split for directly comparable benchmark results.

Citation

Citation information and the BibTeX entry will be updated after the associated paper is published.

% To be updated after publication.

License

PolyMFO is licensed under the CC BY-NC 4.0 license.

Non-commercial use is permitted with proper attribution. Commercial use is prohibited.

See the full license terms: https://creativecommons.org/licenses/by-nc/4.0/