Deep Learning Based Approach to Generate Realistic Data for ADAS Applications
Rajat Kumar Soni, Binoy B. Nair · 2021
Quantity, quality and diversity of datasets are prerequisites for training the deep-learning autonomous driving models. One major issue identified from the literature is the lack of realistic training data which eventually leads to a less robust model. Simulators can help in dealing with reducing reality gaps, however, commonly available simulators generate data that are far removed from the real world scenarios. The model proposed in this study is based on video-to-video synthesis and image synthesis methods using Generative Adversarial Networks (GANs). The results indicate improved realism. Kanade-Lucas-Tomasi (KLT) and Fr'echet Inception Distance (FID) based temporal coherence evaluation metrics have also been proposed as a possible alternative to human perception driven evaluations.