## Dataset Information
This dataset focuses on extracting pitting corrosion within Aluminum (Al) wires using in-situ XCT images acquired at Sandia National Laboratories (SNL).
Currently, the dataset consists of 100 annotated images.
#### **Instructions for Generating Predictions and Evaluating Model Performance on the Benchmark Dataset**
- Images and corresponding groud truth masks are [here][1].
- For each model, we provide an inference script to generate predictions.
- Refer to the table at the bottom for links to the inference scripts for each model.
- You need to change the paths for images and model in the script accordingly.
- Once you've generated the predictions, you can use the '[get_evaluation_metrics.py][2]' script (applicable to all models in this dataset) to assess the model's performance on the generated predictions.
- You need to change the paths for images in the script accordingly.
**Metadata:**
- **Imaging technique:** XCT
- **Sample:** Commercial-purity Aluminum (1100 Al) wire
- **Diameter:** 0.813 mm
- **Length:** 1.25 mm
- **Sample exposure:**
- **NaCl droplets:** 60 µg/cm² NaCl
- **Relative Humidity:** 98%
- **Image size:** 1024 x 988
- **Voxel size:** 1.25 µm
- **Minimum resolvable feature size:** 15.6 µm² (2x2x2 voxels)
**Want to contribute? Fill up this [Google Form][3]**.
<h2>
Benchmarks
</h2>
| Model | Description | Binary IoU | Library | Input Image Size | Performance Metrics | Paper | Model Inference Script | Compatible PyTorch Version |
|:---------------------------------:|:--------------------------------------------:|:----------:|:--------------------:|:----------------:|:------------------------------------------:|:------------------------------------------:|:------------------------------------------:|:------------------------------------------:|
[unet-seresnext101.h5][4]| U-Net with SE-ResNeXt101 Encoder Backbone | 56%|Tensorflow/Keras <br> ([Segmentation Models][5]) | 768 x 768| Accuracy <br> Precision <br> Recall <br> Binary IoU* | - | [Click Here][6] | TensorFlow 2.17.0 |
|[unet-standard.h5][7]| U-Net model with 4-block encoder and decoder| 48% |Tensorflow/Keras | 768 x 768| Accuracy <br> Precision <br> Recall <br> Binary IoU*| - | [Click Here][8] | PyTorch 2.13.0 |
[1]: https://osf.io/uzxpc/files/osfstorage
[2]: https://osf.io/download/xpu4r/
[3]: https://forms.gle/4pGRhkdw4wGNy1dY9
[4]: https://osf.io/download/7mqh8/
[5]: https://github.com/qubvel/segmentation_models
[6]: https://osf.io/download/6707e5ddb6fe7aae902eca72/
[7]: https://osf.io/download/auzts/
[8]: https://osf.io/download/6707e5de3783bf7ee5a5e0ff/
