Cutout and CutMix in Computer Vision - Cheat Sheet & Quick Revision

Recall & Review

beginner

What is the main idea behind the Cutout data augmentation technique?

Cutout randomly masks out a square patch of the input image during training. This forces the model to focus on other parts of the image, improving robustness and generalization.

intermediate

How does CutMix differ from Cutout in data augmentation?

CutMix replaces a random patch of one image with a patch from another image and mixes their labels proportionally. Unlike Cutout which masks with zeros, CutMix combines two images and their labels.

intermediate

Why does CutMix use mixed labels when combining two images?

Because the new image contains parts from two different images, the label is a weighted combination of both original labels. This helps the model learn from mixed features and improves generalization.

beginner

What is a practical benefit of using Cutout or CutMix during training?

Both techniques help the model become less sensitive to specific parts of images and reduce overfitting. This leads to better performance on new, unseen images.

advanced

In Cutout, what happens if the masked patch is too large?

If the masked patch is too large, the model may lose too much information from the image, making learning harder and possibly hurting performance.

What does Cutout do to an image during training?

A Swaps patches between two images

B Masks a random square patch with zeros

C Rotates the image randomly

D Adds noise to the entire image

Cutout masks a random square patch with zeros to force the model to learn from other parts.

How does CutMix combine labels of two images?

A Mixes labels proportionally to the area of patches combined

B Uses only the label of the second image

C Uses only the label of the first image

D Ignores labels during training

CutMix mixes labels proportionally to the area of the patches taken from each image.

Which technique replaces a patch of one image with a patch from another image?

A CutMix

B Cutout

C Random Crop

D Horizontal Flip

CutMix replaces a patch of one image with a patch from another image.

What is a key goal of using Cutout or CutMix in training?

A To increase model size

B To make images grayscale

C To reduce overfitting and improve generalization

D To speed up training time

Both techniques help reduce overfitting and improve the model's ability to generalize.

If the masked patch in Cutout is too large, what might happen?

A Model learns faster

B Model accuracy increases automatically

C Model ignores the patch

D Model loses too much information and performance drops

A too large masked patch removes too much information, making learning harder and hurting performance.

Explain how Cutout and CutMix help improve model training in computer vision.

Think about how hiding or mixing parts of images forces the model to learn better features.

Describe the difference between Cutout and CutMix in terms of image and label manipulation.

Focus on what happens to the image pixels and labels in each method.