SSMask: Salient Super-Pixel Mask Data Augmentation for Object Detection

Data augmentation is a widely used regularization method to improve the robustness of Convolutional Neural Networks (CNNs) by alleviating the over-fitting problems. In this paper, we propose a heuristic data augmentation scheme for object detection task, termed Salient Super-pixel Mask (SSMask), which exploits the saliency to determine the most discriminative information and achieve the balance between information removal and preservation by the number of seeds of super-pixel segmentation. Concretely, we first obtain sub-regions of an input image by the super-pixel segmentation operator. Then, the static saliency detection algorithm is utilized to compute the saliency map. The most discriminative region can be determined by the combination of super-pixel segmentation and saliency map. Lastly, a region removal mask is built by removing the most discriminative super-pixel region. The data augmentation scheme will be randomly applied to the training stage. Extensive experiments are performed on the Pascal VOC dataset. The mean Average Precision (mAP) of RetinaNet is boosted from 81.59% to 83.14%. The mAP of CenterNet is improved from 81.84% to 84.13%, and the mAP of FCOS is boosted from 79.10% to 83.11%. Experimental results prove the effectiveness of the proposed method.