Research on Evaluation Method of Image Blending Based Image Simulation
Yu Zhang, Yan Zhang, Zhiguang Shi, Xiaoran Shi, Di Liu, Yuchang Suo · 2021
Object detection and tracking algorithms for specific targets are important technologies in the fields of autonomous driving and video surveillance. The verification of related algorithms and the training of data-driven machine learning methods often rely on high-quality and large-scale data support. Therefore, data simulation methods are usually used to supplement the quantity and richness of data. This paper proposes an image simulation performance evaluation method based on the Probably Approximately Correct (PAC) learnable theory: experience consistency loss, which is used to quantitatively evaluate simulated images. Compared with the traditional performance evaluation method based on expert score and Turing test, this evaluation method can objectively and quantitatively reflect the characteristic consistency between the simulated image and the real image. The experimental results show that the evaluation results of this method are consistent with the qualitative analysis results, which can directly reflect the performance improvement of the simulation image for the machine learning model.