Underwater real-time object recognition is essential to unmanned underwater drones, namely autonomous underwater vehicles (AUV), cruising in the ocean. As the deep learning technology evolves swiftly lately, the attempt for AUVs to fully understand the surrounding environment becomes an emerging demand for marine or military applications. No matter which approach that deep learning manages to adopt, a large dataset with sufficient number of images for each object is required. In this investigation, a dataset augmentation method based on 3D modeling is proposed to resolve the mentioned problem. By rotating and scaling the target object in 3 dimensions with different backgrounds, the number of underwater object images is increased over 1000 times. Through the proposed method, high quality image data are forged to improve the recognition accuracy of those rare underwater objects, which are very hard to collect enough number of images, by 20% based on real-time video clips' experiment.