Enhancing the Data Regularization Effect with Randomly Combined Features for Object Detection

Deep convolutional neural networks (CNNs) have made significant performance improvements on object detection and several augmentation techniques have been introduced to further improve the detection performance. We investigate how applying multiple augmentation techniques simultaneously can affect the learning capability of existing techniques and how providing more abundant training backgrounds to an image can have an effect. Our experimental results demonstrate that the performance has improved by combining Random Perceptive and Random Erasing techniques to Mosaic techniques on PASCAL VOC dataset. Our results also show that the combination of the augmentation techniques is also effective in small-sized specific datasets.