SmoothMix, soft-edge based data mix augmentation
Data augmentations have long been a big task in machine learning. It is originally a strategy to increase the training data size. But later researches showed that proper data augmentation strategies can increase the robustness for trained models toward adversarial examples.
SmoothMix is a novel augmentation strategy proposed by the paper “SmoothMix: a Simple Yet Effective Data Augmentation to Train Robust Classifiers” accepted by CVPR2020.
Introductions
Recent studies like Cutout, Random erasing suggests remove certain region in data will raise the robustness of the model, and data-mix type augmentations like Cutmix, Mixed example, suggests replacing the area with other images and using multi-label for training.
However, these regional dropout methods may lead to a ‘strong-edge’ problem due to the drastic change in pixels because of the replaced regions.
As its name implies, SmoothMix mix data smoothly. Instead of replacing regions, SmoothMix using a Gaussian base blending mask to mix two images with soft-edge.
How to soften edge?
The way to two images sampled from the training dataset into SmoothMix data is to add after an element-wise multiplication with masks G and 1 − G respectively. The resulting image is a mix of the two input images, smoothly merged into each other.
Why Soft-edge?
This paper investigates the impacts of strong-edge based augmentation strategies by using Class Activation Map(CAM), as shown in the above image.
It’s loud and clear that if part of the original image was replaced by strong-edge shape, the activation map of the certain region will somehow being highly activated, and suppress the activation of the original region.
As the paper implies, although strong-edge based strategies improved over the vanilla model, it gets distracted by the box occlusion and does not show significant improvement in output score.
Results
Although the Top-1 error(%) on every dataset was slightly lower than CutMix, it still outperformed Baseline and provided an alternative way to augment the data.
Conclusions
SmoothMix is indeed a good way to augment the data, especially for those who don’t like strong-edge strategies (like me!). It provides mixed data that don’t look like a jigsaw puzzle, and as results shows, the activation maps seem a lot more smooth.
if you like(this_article):
please(CLAPS)
follow(LSC_PSD)
# Thanks :)