Generating Realistic Images with NeRF for Training of Autonomous Driving Network
Sugaya Shun, Yuichi Okuyama, Shintomi Yuta, Kusano Ryota, Tanaka Kohsuke, Shin Jungpil, Satoshi Nishimura · 2023
We present method of constructing dataset of training NeRF for synthesizing a realistic autonomous training dataset. Since the realism of NeRF depends on the image set used for training, it is necessary to use an appropriate image set. Therefore, to generate realistic synthesized images in the autonomous vehicle racing environment, we aim to improve the realism with images from in-vehicle viewpoints in addition to images obtained by the photography technique widely used in the research on NeRF. We experiment with a 1/10 scale vehicle to evaluate the driving performance of the autonomous driving models trained with synthesized images by NeRF. In this experiment, we compare the model’s driving performance using our proposed NeRF training image set with an image set obtained by a widely used photography technique. The experimental results show that the autonomous driving model trained with our proposed method drives better on a real racetrack than the model trained with the conventional method.