Re-Labeling for Real-time Semantic Segmentation in Specific Environments

In this paper, we perform re-labeling to evaluate the performance of semantic segmentation in a specific environment. By verifying the performance given new type of labels, the potential for future applications is confirmed. To check the performance, we use the BiSeNet V2 network which shows good performance in real-time semantic segmentation. In BiSeNet V2, low-level details and high-level semantics are individually trained to achieve high accuracy and real-time semantic segmentation performance. Using the reconstructed labels from the Cityscapes dataset composed of 2048x1024 images, the real-time semantic segmentation performance was evaluated, achieving an accuracy of 80.9% and a speed of 111 FPS.