Data Augmentation on Synthetic Images for Transfer Learning using Deep CNNs

Training of deep Convolutional Neural Networks (CNNs) for object detection tasks requires a huge amount of annotated data which is expensive, difficult and time-consuming to produce. This requirement can be fulfilled by automating the process of dataset generation. We utilize the approach of training deep CNNs using completely synthetically rendered data, with the focus of improving the overall transfer learning performance through online and offline data augmentation techniques. We focus on the problem of detecting packaged food products in indoor refrigerator environments. We analyze the impact of various data augmentation strategies like randomized cropping, pixel shifting, image scaling, image rotation, oversaturation, Gaussian blurring, noise addition, color inversion etc. on the overall accuracy of the object detection and increase the overall mean average precision (mAP). It is found that the use of a combination of data augmentation techniques performs best, with highest mAP of 20.54 obtained with combinations of linear augmentation techniques like scaling, shifting and scaling and rotation.