Training Deep Neural Networks with Synthetic Data for Off-Road Vehicle Detection

In tandem with growing deep learning technology, vehicle detection using convolutional neural network is now become a mainstream in the field of autonomous driving and ADAS. Taking advantage of this, lots of real image datasets have been produced in spite of the painstaking work of data collection and ground truth annotation. As an alternative, virtually generated images are introduced. This makes data collection and annotation much easier, but a different kind of problem called `domain gap' is announced. For instance, in off-road vehicle detection, there is a difficulty in producing off-road image dataset not only by collecting real images, but also by synthesizing images sidestepping the domain gap. In this paper, focusing on the off-road army tank detection, we introduce a synthetic image generator using domain randomization on off-road scene context. We train a deep learning model on synthetic dataset using low level features form feature extractor pre-trained on real common object dataset. With proposed method, we improve the model accuracy to 0.86 AP@0.5IOU, outperforming naïve domain randomization approach.