| Summary: | 碩士 === 國立中興大學 === 資訊科學與工程學系所 === 107 === In this thesis, we validate the idea of using 3D models to generate 2D image dataset for training object detection model as a mechanism of data augmentation. The idea is to rotate 3D models to generate images of different angles and synthesize the images with texture to enhance the quality of data augmentation. In the experiment, we select aircraft and tank as target objects. The training results found that: using the aircraft as the target object, the accuracy of the model trained by real images is only 51.9%, which is lower than the result of the training with our data augmentation method (79%). As a result of using the tank as the target object, it is also verified that the accuracy of our data augmentation method (72.9%) is higher than that of the training with real images (51.3%). The experiment result validates show that this data augmentation technique provides comparable performance with respect to the object detection model trained with the same number of real images.
|
|---|
