FMix

FMix is a data augmentation technique used in training deep learning models, particularly for image classification. It blends two training samples by applying a randomly generated mixing mask, producing smooth, irregular regions of overlap rather than sharp cuts. The goal is to encourage the model to rely on distributed evidence and to improve generalization, while reducing artifacts that can arise from other mixing methods.

Mask generation and mixing process: A random field is constructed in Fourier space and filtered to emphasize low-frequency components, yielding a smooth spatial mask after an inverse transform. The area of the mask is controlled by sampling a mixing coefficient from a Beta distribution, which determines how much of the second image will contribute to the mix. The resulting mask is then used to combine two images: the mixed input is x' = x1 * (1 - M) + x2 * M. The corresponding label is also mixed using the same proportion, typically y' = y1 * (1 - lambda) + y2 * lambda, where lambda reflects the mask’s average value.

Relation to other methods and usage: FMix is one of several mix-based augmentation strategies that build on ideas from Mixup and CutMix. It is used to increase data diversity during training and can be adapted with different alpha values and masking parameters to control the smoothness of edges and the expected mixing ratio. While often beneficial for vision tasks, FMix may require tuning for specific datasets and can introduce additional computational overhead due to the Fourier-domain operations.