Data augmentation to improve robustness of image captioning solutions

TL;DR

This paper explores how data augmentation can enhance the robustness of image captioning systems against motion blur, significantly reducing performance degradation on benchmark datasets.

Contribution

It introduces a data augmentation approach applied at both object detection and captioning stages to improve robustness to motion blur in image captioning.

Findings

Augmenting both stages reduces CIDEr-D degradation significantly.

Performance degradation decreases from 68.7 to 11.7 on MS COCO.

Performance degradation decreases from 22.4 to 6.8 on Vizwiz.

Abstract

In this paper, we study the impact of motion blur, a common quality flaw in real world images, on a state-of-the-art two-stage image captioning solution, and notice a degradation in solution performance as blur intensity increases. We investigate techniques to improve the robustness of the solution to motion blur using training data augmentation at each or both stages of the solution, i.e., object detection and captioning, and observe improved results. In particular, augmenting both the stages reduces the CIDEr-D degradation for high motion blur intensity from 68.7 to 11.7 on MS COCO dataset, and from 22.4 to 6.8 on Vizwiz dataset.