Valid Data Augmentation by Patch Alpha Matting
Hongyun Li, Jianghao Rao, Lijun Zhou, Jianlin Zhang · 2019
In this study, we designed a new data augmentation method by matting for object detection and segmentation tasks. This method first uses a trained model to retrieve easy samples from the training data. After the easy dataset is built, a new patch matting algorithm is used to obtain object's fine border, and then a set of photo montage strategies is used to generate a new dataset. For natural image matting, obtaining accurate borders around very similarly colored foreground and background regions is a hard challenge. This problem is often encountered in reality and causes visual flaws in composite images. To avoid poor composite images, we proposed a new matting method based on color and semantic features on patches to obtain visually acceptable images in such challenging regions and used it in the data augmentation method above. Unlike the traditional data augmentation method, which can only generate similar images, our method can effectively improve the ability of models to distinguish objects from the background. In experiments, we generated 20,000 new images and added them to the original COCO dataset to train a MaskRCNN model. As a result, the performances of our model are superior to the original model on all evaluating indexes of COCO.