[Paper Review] GridMask Data Augmentation

Pengguang Chen, Shu Liu+3 more|arXiv (Cornell University)|Jan 13, 2020

Advanced Neural Network Applications35 references226 citations

TL;DR

GridMask is a simple, structured information-dropping augmentation that improves CNN performance across image classification, object detection, and semantic segmentation, outperforming prior dropping methods and AutoAugment with lower computation.

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.

Motivation & Objective

• Motivate effective data augmentation by balancing deletion and preservation of image regions.
• Propose a simple grid-based masking strategy that drops information in a structured, non-contiguous way.
• Demonstrate GridMask’s effectiveness across multiple vision tasks and datasets.
• Analyze parameter choices and variations to establish robust guidelines for practice.

Proposed method

• Represent masking as a binary grid mask M applied to input x via x~ = x * M.
• Define grid units with parameters (r, d, delta_x, delta_y) to control keep ratio k and dropped region size l = r * d.
• Empirically study keep ratio, unit size, and displacement ranges to minimize unsuccessful augmentation cases.
• Rotate masks and vary d during training to increase augmentation diversity.
• Evaluate GridMask on ImageNet, COCO, and Cityscapes against baselines and AutoAugment, including ablation variations such as reversed GridMask and random dropping.

Experimental results

Research questions

• RQ1 Does a structured grid-based dropping strategy improve generalization across classification, detection, and segmentation tasks?
• RQ2 How do mask parameters (r, d, delta_x, delta_y) affect performance and robustness?
• RQ3 Can GridMask outperform existing information dropping methods and AutoAugment with lower computational cost?
• RQ4 What are the effects of grid mask variations (reversed GridMask, random dropping) on results?

Key findings

• GridMask improves ImageNet ResNet50 from 76.5% to 77.9% and ResNet152 from 78.3% to 79.7%.
• GridMask improves COCO2017 Faster-RCNN-50-FPN mAP from 37.4% to 39.2% and Faster-RCNN-X101-FPN from 41.2% to 42.6%.
• GridMask improves Cityscapes PSPNet50 mIoU from 77.3% to 78.1%.
• On CIFAR-10, GridMask yields higher accuracy across several models (e.g., ResNet-18 from 95.28% to 96.54%).
• GridMask outperforms Cutout, HaS, and AutoAugment on several settings, while adding minimal computational overhead.
• A reversed GridMask and structured dropping outperform random dropping, demonstrating robustness of the approach.