[Paper Review] Data Augmentation by Pairing Samples for Images Classification

Hiroshi Inoue|arXiv (Cornell University)|Jan 9, 2018

Machine Learning and Data Classification11 references338 citations

TL;DR

SamplePairing overlays two randomly chosen training images to synthesize a new training sample, using the first image's label, yielding significant accuracy gains across multiple datasets, especially with limited data.

ABSTRACT

Data augmentation is a widely used technique in many machine learning tasks, such as image classification, to virtually enlarge the training dataset size and avoid overfitting. Traditional data augmentation techniques for image classification tasks create new samples from the original training data by, for example, flipping, distorting, adding a small amount of noise to, or cropping a patch from an original image. In this paper, we introduce a simple but surprisingly effective data augmentation technique for image classification tasks. With our technique, named SamplePairing, we synthesize a new sample from one image by overlaying another image randomly chosen from the training data (i.e., taking an average of two images for each pixel). By using two images randomly selected from the training set, we can generate $N^2$ new samples from $N$ training samples. This simple data augmentation technique significantly improved classification accuracy for all the tested datasets; for example, the top-1 error rate was reduced from 33.5% to 29.0% for the ILSVRC 2012 dataset with GoogLeNet and from 8.22% to 6.93% in the CIFAR-10 dataset. We also show that our SamplePairing technique largely improved accuracy when the number of samples in the training set was very small. Therefore, our technique is more valuable for tasks with a limited amount of training data, such as medical imaging tasks.

Motivation & Objective

• Motivate improving generalization in image classification with limited training data.
• Introduce a simple data augmentation method that enlarges the effective training set by pairing samples.
• Evaluate SamplePairing across standard benchmarks and analyze its interaction with other augmentations.

Proposed method

• Synthesize a new image by averaging two randomly chosen training images per training instance.
• Assign the label of the first image to the mixed image, ignoring the second image's label.
• Integrate SamplePairing with existing augmentations without modifying network architectures.
• Evaluate under staged training where SamplePairing is enabled after initial epochs and periodically disabled during training as fine-tuning.
• Compare with related methods (e.g., mixup, Between-class learning) to isolate benefits of equal-weight pairing without label blending.

Experimental results

Research questions

• RQ1 Does SamplePairing improve generalization across datasets (ILSVRC, CIFAR-10/100, SVHN) compared to baseline augmentations?
• RQ2 How does selecting the overlay image from training data versus outside data affect performance?
• RQ3 Is equal-weight averaging (no label blending) as effective as methods that blend labels or use weighted mixtures?
• RQ4 How does SamplePairing perform when the training set is small or imbalanced?
• RQ5 What is the impact of intermittently disabling SamplePairing during training on final performance?

Key findings

• SamplePairing reduces validation error on ILSVRC, CIFAR-10, CIFAR-100, and SVHN compared to baselines.
• The top-1 error on ILSVRC with GoogLeNet dropped from 33.5% to 29.0%.
• CIFAR-10 error dropped from 8.22% to 6.93% with a simple 6-layer network.
• With restricted training data (e.g., 100 samples per class in CIFAR-10), accuracy gains are larger (e.g., 43.1% to 31.0%).
• Overlaying an image from outside the training set yields improvements but is less effective than using an in-training-set overlay.
• Using the label from only one image (no label blending) yields performance comparable to methods that mix labels.