Benchmarking Deep Learning Models for Instance Segmentation

Instance segmentation has gained attention in various computer vision fields, such as autonomous driving, drone control, and sports analysis. Recently, many successful models have been developed, which can be classified into two categories: accuracy- speed-focused. Accuracy inference time are important for real-time applications of this task. However, these just present measured on different hardware, makes their comparison difficult. This study is the first to evaluate compare performances state-of-the-art instance by focusing a fixed experimental environment. For precise comparison, test hardware environment should identical; hence, we accuracy speed quantitative qualitative analyses. Although speed-focused run high-end GPUs, there trade-off between when computing power insufficient. The results show that feature pyramid network structure may considered designing model, balance must achieved application.