fmix

FMix is a data augmentation technique that generates mixing masks using Fourier-space filtering — creating random, irregular binary masks with controllable frequency characteristics for more diverse and natural-looking image combinations.

How Does FMix Work?

• Generate Noise: Create random noise in Fourier space with controlled frequency spectrum.
• Filter: Low-pass filter to control the smoothness of the resulting mask.
• Threshold: Convert to binary mask via thresholding.
• Mix: Apply the mask to combine two images, with labels mixed proportionally to mask area.
• Paper: Harris et al. (2020).

Why It Matters

• Diverse Masks: Fourier-generated masks produce irregular, organic shapes unlike CutMix's rectangles.
• Controllable: The frequency cutoff controls mask complexity (smooth blobs vs. fine-grained patterns).
• Better Than CutMix: Irregular masks prevent the model from exploiting rectangular boundary patterns.

FMix is Fourier-designed masks for mixing — using frequency-domain generation to create diverse, naturalistic mixing patterns.