GridMask Data Augmentation

TL;DR

GridMask is a new data augmentation technique that systematically removes regions of input images, leading to improved performance across various computer vision tasks with less computational cost than existing methods.

Contribution

The paper introduces GridMask, a simple yet effective structured information removal method that outperforms state-of-the-art augmentation techniques like AutoAugment.

Findings

Outperforms AutoAugment on multiple datasets

Improves recognition, detection, and segmentation performance

Less computationally expensive than reinforcement learning-based methods

Abstract

We propose a novel data augmentation method `GridMask' in this paper. It utilizes information removal to achieve state-of-the-art results in a variety of computer vision tasks. We analyze the requirement of information dropping. Then we show limitation of existing information dropping algorithms and propose our structured method, which is simple and yet very effective. It is based on the deletion of regions of the input image. Our extensive experiments show that our method outperforms the latest AutoAugment, which is way more computationally expensive due to the use of reinforcement learning to find the best policies. On the ImageNet dataset for recognition, COCO2017 object detection, and on Cityscapes dataset for semantic segmentation, our method all notably improves performance over baselines. The extensive experiments manifest the effectiveness and generality of the new method.