CutMix

CutMix is a data augmentation technique used in training deep neural networks for computer vision. It creates new training samples by taking a patch from one image and pasting it into another, and blending the corresponding labels according to the patch area. The method aims to encourage the model to attend to multiple regions of an image and to improve generalization.

Procedure: Given two samples (x1, y1) and (x2, y2), a mixing ratio lambda is drawn from a Beta(alpha, alpha) distribution. A rectangular patch is selected with area proportional to 1 − lambda and is placed into the other image at a random position. The resulting image is x′ with label y′ = lambda*y1 + (1 − lambda)*y2 (soft labels).

Variations and parameters: Different implementations vary the shape and size of the pasted patch and the distribution used to sample lambda. The alpha parameter controls the strength of mixing; typical choices range from 0.3 to 1.0. Some approaches combine CutMix with other augmentation techniques such as Mixup or Cutout, or adapt patch placement to align with salient image regions.

Applications and effects: CutMix has been shown to improve accuracy and robustness on image classification benchmarks by regularizing models and reducing overfitting. It can be applied to other visual tasks, including object detection and segmentation, though effects on localization signals may be task-dependent and require careful tuning.

Implementation notes: CutMix is relatively simple to implement and is supported in many deep-learning frameworks. When using mixed labels, losses are computed against the combined target. Hyperparameters, especially alpha, should be chosen based on the dataset and task.