#### NOTE
This is a **Hugging Face dataset**. For large datasets, ensure `huggingface_hub>=1.1.3` to avoid rate limits. Learn more in the Hugging Face integration docs.
# Dataset Card for Kolektor Surface-Defect Dataset
KolektorSDD (Kolektor Surface-Defect Dataset) is a grayscale industrial surface-inspection
dataset of electrical commutators.
This FiftyOne dataset uses the **box-annotation release** intended for the ICPR 2021 and
COMIND 2021 papers ([download](https://go.vicos.si/kolektorsddboxes)): one sample per
surface image, with defect regions annotated as axis-aligned bounding boxes stored as filled
rectangles in the label masks.
This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 399 samples.
## Installation
If you haven’t already, install FiftyOne:
```bash
pip install -U fiftyone
```
## Usage
```python
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/Kolektor_Surface_Defect")
# Launch the App
session = fo.launch_app(dataset)
```
## Dataset Details
- **Curated by:** Domen Tabernik, Samo Šela, Jure Skvarč, Danijel Skočaj (University of Ljubljana / ViCoS Lab); images provided and annotated by Kolektor Group d.o.o.
- **Paper (dataset):** [Segmentation-Based Deep-Learning Approach for Surface-Defect Detection](https://doi.org/10.1007/s10845-019-01476-x)
- **Box annotations used in:** [End-to-end training of a two-stage neural network for defect detection](https://arxiv.org/abs/2007.07676) (ICPR 2020) and [Mixed supervision for surface-defect detection](http://prints.vicos.si/publications/385) (Computers in Industry, 2021)
- **Project page:** https://www.vicos.si/resources/kolektorsdd/
- **Download (this release):** https://go.vicos.si/kolektorsddboxes
- **License:** [CC BY-NC-SA 4.0](http://creativecommons.org/licenses/by-nc-sa/4.0/) (non-commercial; contact [Danijel Skočaj](https://www.vicos.si/people/danijel_skocaj/) for commercial use)
### What the data contains
Images were captured in a controlled industrial environment. Each sample is one
non-overlapping view of a commutator surface. Defects are microscopic fractures or
cracks in the plastic embedding.
| Property | Value |
|-------------------------|-------------------------------------|
| Total images | 399 |
| Physical items (boards) | 50 (`kos01`–`kos50`) |
| Surfaces per item | 8 (`Part0`–`Part7`) |
| Defective images | 52 |
| Non-defective images | 347 |
| Image type | Grayscale JPG |
| Original size | 500 px wide × 1240–1270 px tall |
| Recommended eval size | 512 × 1408 px (per dataset authors) |
Defect visibility: for 48 items the defect appears in exactly one image; for 2 items
it appears in two images.
A separate **fine pixel-annotation release** exists for the JIM2019 paper
([download](https://go.vicos.si/kolektorsdd)). That version is not what this card describes.
### Raw download layout
```default
kolektorsdd/
kos01/
Part0.jpg
Part0_label.bmp
Part1.jpg
Part1_label.bmp
...
kos02/
...
```
- `Part*.jpg` — surface image
- `Part*_label.bmp` — defect annotation mask (non-zero = defect region)
In this box-annotation release, each defective mask is a **filled axis-aligned bounding box**
around the defect, not a precise pixel-wise segmentation of the crack shape.
### Train/test splits
The authors evaluate with **3-fold cross-validation**, keeping all 8 images of the
same physical item in the same fold. Official split files:
[KolektorSDD-training-splits.zip](https://data.vicos.si/datasets/KSDD/KolektorSDD-training-splits.zip).
This FiftyOne dataset does **not** assign fold/split labels. Add them externally if needed.
---
## FiftyOne Dataset Structure
| Property | Value |
|--------------------|------------------------------------------|
| Hub dataset | `harpreetsahota/Kolektor_Surface_Defect` |
| Local dataset name | `kolektorsdd` |
| Media type | `image` |
| Samples | 399 |
### Sample fields
| Field | Type | Description |
|----------------|---------------------------------------|--------------------------------------------------|
| `filepath` | `StringField` | Path to source `Part*.jpg` |
| `board_id` | `StringField` | Board directory name, e.g. `"kos01"` |
| `has_defect` | `BooleanField` | `True` if the mask contains any foreground pixel |
| `ground_truth` | `EmbeddedDocumentField(Segmentation)` | Binarized mask (`0` = background, `1` = defect) |
The local parser (`parse_to_fo.py`) reads each BMP label and stores a `{0, 1}` mask on
the sample. For defective images in this release, the foreground region is a filled
bounding box rather than a tight defect outline.
---
## Citation
**BibTeX (dataset):**
```bibtex
@article{Tabernik2019JIM,
author = {Tabernik, Domen and {\v{S}}ela, Samo and Skvar{\v{c}}, Jure and Sko{\v{c}}aj, Danijel},
journal = {Journal of Intelligent Manufacturing},
title = {{Segmentation-Based Deep-Learning Approach for Surface-Defect Detection}},
year = {2019},
month = {May},
day = {15},
issn = {1572-8145},
doi = {10.1007/s10845-019-01476-x}
}
```
**BibTeX (box annotations / mixed supervision):**
```bibtex
@article{Bozic2021COMIND,
author = {Bo{\v{z}}i{\v{c}}, Jakob and Tabernik, Domen and Sko{\v{c}}aj, Danijel},
journal = {Computers in Industry},
title = {{Mixed supervision for surface-defect detection: from weakly to fully supervised learning}},
year = {2021}
}
```
**APA:**
Tabernik, D., Šela, S., Skvarč, J., & Skočaj, D. (2019). Segmentation-Based Deep-Learning Approach for Surface-Defect Detection. *Journal of Intelligent Manufacturing*. https://doi.org/10.1007/s10845-019-01476-x