[LS2N_IPI]Concrete patch classification

concrete_patch_classification

This dataset is based on the original dataset I3DCP introduced in Rill-García, R., Dokladalova, E., Dokládal, P., Caron, J.-F., Mesnil, R., Margerit, P., & Charrier, M. (2022). Inline monitoring of 3D concrete printing using computer vision. Additive Manufacturing, 60, 103175. https://doi.org/10.1016/j.addma.2022.103175

The original dataset includes raw images of cement-based material deposition, segmentation masks of interstitial lines, and texture classification patches. In particular, our work focuses on the texture classification patches.  This dataset thus provides three complementary resources:

A reorganized version of the original 111 patches with 5-fold splits.

An extended set of 426 expert-annotated patches with an additional geometric defect class(Crushed in English, Écrasé in French).

A collection of synthetic patches generated with StyleGAN3, covering all five classes.

Sub-dataset 1: Original annotated texture windows

Content: 111 labeled gray-leveled texture windows with fixed width 200 extracted from 24 raw images. 5-fold cross-validation

Original classes:

Fluid (24 images, proportion 21.62%)

Good (27 images, proportion 24.32%)

Dry (24 images, proportion 21.62%)

Tearing (36 images, proportion 32.43%)

Labels: texture_windows-labels.csv.

Model weights fine-tuned in subdataset1 with synthetic images in subdataset3: Baseline model introduced by (Rill-García et al., 2022) , EfficientFormer model introduced by (Li et al., 2022) and proposed Multimodal Dual-Branch model. (pth: model weight, *.txt: normalization params for image, *.npy: normalization params for texture descripteur vector) 

Sub-dataset 2: Extended expert-annotated texture windows

Content: 426 extended labeled gray-leveled texture windows with fixed width 200 extracted from 24 raw images. 5-fold cross-validation

Classes:

Fluid(84 images,proportion 19.72%)

Good(127 images,proportion 29.81%)

Dry(68 images,proportion 15.96%)

Tearing(61 images,proportion 14.32%)

Geometric defect Écrasé (French) / Crushed (English) (86 images, proportion 20.19%)

Labels: patch_labels(426extension).csv

Model weights fine-tuned in subdataset2 with synthetic images in subdataset3: Baseline model introduced by (Rill-García et al., 2022) , EfficientFormer model introduced by (Li et al., 2022) and proposed Multimodal Dual-Branch model.(pth: model weight, *.txt: normalization params for image, *.npy: normalization params for texture descripteur vector) 

Sub-dataset 3: Synthetic images (StyleGAN3 generated)

Content: Synthetic gray-leveled texture windows generated by five separate pretrained generative models.

Classes:

Fluid(1200 images)

Good(1200 images)

Dry(1200 images)

Tearing(1200 images)

Geometric defect Écrasé (French) / Crushed (English)(1200 images)

Labels: ./images_generees(d1)/patch_labels(426extension+stylegan3).csv  for Sub-dataset2.   ./images_generees(d2)/texture_windows-labels(stylegan3_d2).csv  for Sub-dataset1.

Model weights trained for generation: 4 category-specific model weights trained by StyleGAN3 (fluid, good, dry, tearing), each model can only generate one category. 1 category-jointly model weights trained by StyleGAN3, which generates 5 categories(fluid, good, dry ,tearing, ecrase/crushed)

For specific dataset usage, please refer to the GitHub repository

Updates (compared to Version 1.0.0)

The models were re-trained under an updated training configuration, resulting in reduced overfitting compared to Version 1.0.0. In addition, the inference procedure has been upgraded from a single-model setup to a 5-model ensemble strategy based on logits averaging. 

Synthetic Data Extension (SubDataset3)

The synthetic image dataset has been expanded. In Version 1.0.0, synthetic images were generated exclusively using a generator trained on Original dataset. In the current version, additional synthetic images generated by a generator trained on Re-annotated dataset have been included. The corresponding label CSV files are also provided to facilitate data augmentation during training.

To avoid data leakage between datasets, a cross-dataset generation strategy is adopted:

Synthetic images generated by the generator trained on the Original Dataset (Dataset1) are used exclusively for augmentation of the Re-annotated Dataset (Dataset2).

Conversely, synthetic images generated by the generator trained on the Re-annotated Dataset (Dataset2) are used exclusively for augmentation of the Original Dataset (Dataset1).

License

This dataset is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).  https://creativecommons.org/licenses/by-nc-sa/4.0/

It is derived from the I3DCP released under the same license (CC BY-NC-SA 4.0). Additional annotations and processing were created by us and are released under the same CC BY-NC-SA 4.0 license.