# Steel Plates Dual Dataset
Visual surface inspection images of steel plates with defects and impurities present.
Dataset was used to support study of influence of impurities on anomaly detection.
Plates were inspected using a ring light and grayscale camera monted on a robotic manipulator.
The dataset is a collection of grayscale PNG image files easily readable by any software.

The dataset is supporting the publication:
```
Sequential PatchCore: Anomaly Detection for Surface Inspection using Synthetic Impurities
Runzhou Mao, Juraj Fulir, Christoph Garth, Petra Gospodnetić
2nd workshop on Vision-based InduStrial InspectiON (VISION), ECCV 2024
Milan, Italy
```

## Subsets
The dataset contains 8 subsets which are used for various purposes.

- `Real_defected` - Contains images from multiple viewpoints of real steel plates containing defects and the only subset containing corresponding defect masks.

  * Used only for threshold estimation and testing.
  * Defines 4 meaningful k-folds:

      0. `plate_5_a` is the threshold estimation object (has defects and impurities), rest are testing objects. (Used in publication)
      1. `plate_6_a` is the threshold estimation object (has defects and impurities), rest are testing objects.
      2. `plate_5_b` is the threshold estimation object (has only defects), rest are testing objects.
      3. `plate_6_b` is the threshold estimation object (has only defects), rest are testing objects.

- `Real_clean` - Contains images from multiple viewpoints of clean real steel plates
- `Synth` - Contains images from single viewpoint of synthetic clean plates
- `Synth_WS` - Contains images from single viewpoint of synthetic plates with synthetic waterstains
- `Synth_90deglight` - Contains images from single viewpoint of synthetic clean plates augmented by 90 degree rotation of the original hexagonal light source.
- `Synth_WS_90deglight` - Contains images from single viewpoint of synthetic plates with synthetic waterstains augmented by 90 degree rotation of the original hexagonal light source.
- `Synth_biglight` - Contains images from single viewpoint of synthetic clean plates augmented by doubling the size of the original hexagonal light source.
- `Synth_WS_biglight` - Contains images from single viewpoint of synthetic plates with synthetic waterstains augmented by doubling the size of the original hexagonal light source.


## Data structure

```
<subset_root>
|-- images
    |-- <object_name>
        |-- Image_1.png
        |-- ...
|-- segmentation_masks
    |-- <object_name>
        |-- Image_1.png
        |-- ...
|-- instance_masks
    |-- <object_name>
        |-- Image_1.png
        |-- ...
```

All of the subsets share the same structure.
They contain the images and, if applicable, their corresponding segmentation and instance masks sharing the same name and relative path as the image.
The masks are index-based, meaning the pixel integer value corresponds to the class index.


## Classes
The dataset annotates the following classes (listed with corresponding indices):

0. Background
1. Bump
2. Dent
3. Fingerprint
4. Scratch
5. Sticker
6. Waterstain