How many appropriate added synthetic data to improve semantic segmentation?
Punyapat Areerob, Chanon Khongprasongsiri, Vanee Prikboonchan, Disorn Thampithakpong · 2023
Nowadays, autonomous vehicle technology plays a more important role in our lives. Many people pay attention to driverless cars, but, in fact, every industry needs technology that increases safety. But the problem is that the data must be diverse. We proposed a data synthesis in which the number of the two sources needed to be a reasonable ratio. In this paper, the ratio between synthetic and real data is observed experimentally with clearly and noise injection data in the Yolact and Detectron2 model. In the benchmark test, the optimal ratio is 3.5 and 2.5 for Yolact and Detectron2, respectively. Moreover, provided that training data with mixing data method provide better accuracy than one data type training method.