On a Domain Adaptation Method Based on Image Simplification for Machine Learning with Syntheti Data

Abstract

Constructing an object detector by machine learning needs learning data in various situations. Collecting real data spends monetary and human costs. One of the solutions is generating synthetic data based on a simulation. A reproducibility of synthesized data for real scenes affects an accuracy. Synthetic data need to resemble real data. We propose a domain adaptation method to improve an accuracy efficiently. The method applies a filter to both synthetic and real data. The experiments show that simplification by color reduction is effective.