3D fabric scanning for machine learning

The physical properties of fabrics are determined by a wide variety of fibers, and advances in materials science mean that the number of possible textile types is nearly infinite. While various measurement standards have been established, a unified description remains a long way off. This is due to the sheer variety of materials and the complexities of measurement methods. Using deep learning, we decided to train a model to predict the attributes of the KES measurement method, so we first had to normalize the model.

The 3D data is recorded as a point cloud, and the geometry is generated using Poisson reconstruction, so the topology is not uniform. Since different data sizes are harmful to training, we must first unify the data. However, the key to aligning and transferring a uniform topology is to preserve and interpolate as much high-quality geometry as possible.

The process leading up to a model suitable for training:

Types of bad geometry:

Problematic mesh types created by the Poisson reconstruction method. Dimensionality reduction alignment will not work properly unless these are detected and removed, so we implemented a detection method.

• tunnel

• Overlapping

• Perforated

From left to right: tunnel, overlapping, and perforated

Conclusion:

If the original scan model was too bad (wrinkle ridge defects), interpolation is meaningless, so we simply exclude it. Since we were able to reconstruct the fabric 3D model with a probability of just over 97%, we can expand the dataset for model training with a fairly high probability of combining 2D and 3D data.