Object-Based Perspective Transformation Data Augmentation for Object Detection

Abstract

Perspective transformation can mimic the perspective phenomenon in captured images, while few researches focused on utilizing perspective transformation to augment image data. To well exploit its potential in data augmentation, a novel object-based perspective transformation data augmentation framework is proposed in this paper. First, this framework automatically cuts out objects from images based on bounding boxes provided by corresponding annotations. Next, a simulated random perspective transformation is performed on only those cut objects rather than the whole images. The final step consists of the combination of transformed objects and backgrounds, together with calculation of new annotation. Our framework can generate new images by mimicking the perspective phenomenon caused by different orientations of objects and can also introduce new backgrounds at the same time, hence the further augmentation towards insufficient or imbalanced image data without requirements of additional manual manipulations. Experiments are carried out on the Pascal VOC2007 datasets and the results show the effectiveness of our framework on insufficient or imbalanced datasets.