Attentive Mix: An Efficient Data Augmentation Method for Object Detection
Runji Liu, Ying Chen, Jiasheng Wang, Zhaojin Guo · 2021 7th International Conference on Computer and Communications (ICCC) · 2021
Data augmentation is an import method to improve the model performance, many different data augmentation algorithms have been proposed in the field of machine learning, such as Mixup, Cutmix, Cutout, Mosaic, Attentive Cutmix, Dropout, DropBlock. The above algorithms are mainly applied in image classification tasks, but there is still lack of a data augmentation algorithm that can be used directly for object detection tasks. We propose an efficient data augmentation algorithm specifically applied in the field of object detection. Specifically, we select two images from dataset randomly, then use a trained model to generate the attention map from an image and paste it into another image to generate a new image. Finally, we train a new model on the original data with new generated data. The advantage is that it does not lead to discontinuities in the bounding box when applied to object detection and does not lead to overly complex target labels in the newly generated images. We conducted a series of experiments including training by taking a subset of the training set and testing on the full test set to compare the improvement in model robustness of different data augmentation method. Experimental results proof that proposed method achieves about 4% performance improvement on 10%, 50% and 80% fraction of the dataset, which demonstrates that our method can effectively improve the accuracy of the object detection model and enhance the generalization.