Data augmentation has become an essential technique in training deep neural network models to overcome limitations such as overfitting, reduced robustness, and lower generalization. Methods using 3D datasets are among the most common to use data augmentation techniques. However, these techniques are often applied to the entire object, ignoring the object’s local geometry. This is where PatchAugment comes in.
What is PatchAugment?
PatchAugment is a data augmentation framework that applies different augmentation techniques to the local neighborhoods on an object's surface. This is significant because different local neighborhoods on an object can hold a varying amount of geometric complexity. When the same data augmentation techniques are applied at the object level, it is less effective in augmenting local neighborhoods with complex structures. Thus, applying different augmentation techniques to local neighborhoods helps to increase the effectiveness of data augmentation.
How does PatchAugment work?
PatchAugment works by dividing an object's surface into different local neighborhoods or patches. Each patch is then subjected to its own specific data augmentation technique. This means that each part of an object is uniquely transformed to account for its unique geometric features. For example, one patch may undergo rotations, while another patch undergoes scaling. By doing so, PatchAugment can enhance the overall effectiveness of data augmentation, resulting in better classification results.
Experimental Studies
Experimental studies were conducted on the PointNet++ and DGCNN models using four benchmark datasets: ModelNet40 (synthetic), ModelNet10 (synthetic), SHREC’16 (synthetic), and ScanObjectNN (real-world). The results demonstrated that PatchAugment significantly improved the classification accuracy of these models.
In summary, PatchAugment is a groundbreaking data augmentation technique that applies different augmentation techniques to local neighborhoods on an object's surface, resulting in better classification accuracy.
