Prior distributions-based data augmentation for object detection

Abstract

To solve the problem of data hungry, data augmentation methods based on cut-and-paste that can explore the visual context are widely used. However, these methods either limit the expansion of the instances diversity of dataset or increase the computational burden. In this paper, we propose a novel data augmentation strategy based on prior distributions, which can be used to guide data augmentation for object detection. On one hand, the method can effectively capture the relationship between the foreground instance and the visual context. On the other hand, it can increase the instances diversity of the original dataset as much as possible. Experimental results show that the performance of the popular object detection model can be effectively improved by expanding the original dataset with our method. Compared with the baseline, our method improves by 0.8 percentage point on PASCAL VOC and 1.1 percentage points higher on cross-data test set.