Design of a High-Performance Instance Segmentation Model Using Uncertainty-Guided Non-Maximum Suppression and Normalization

Instance segmentation, which is utilized in various applications including autonomous driving, requires a high level of reliability. To meet these requirements, various studies have been conducted on predicting model uncertainty and employing it in the detection and segmentation process. However, existing studies have limitations in that they are susceptible to outliers because they utilize the predicted uncertainty in the detection and segmentation process without additional tuning. In this paper, we propose a robust instance segmentation model that efficiently tackles noise by leveraging normalized uncertainty in post-processing. Furthermore, we propose a novel technique called multi-uncertainty non-maximum suppression, which selects results with higher reliability compared to existing models. The experimental results show that the proposed method outperforms the existing baseline by achieving a performance improvement of 2.9%.