Benchmarking lightweight deep learning models on real-time semantic image segmentation

Abstract

Real-time semantic image segmentation is critical for applications in autonomous systems, robotics, and edge computing, where computational efficiency and accuracy must be balanced. This study systematically evaluates various lightweight deep learning architectures for semantic segmentation using a comprehensive benchmarking framework on a standardized environment with the Nvidia V100 GPU. We assess models across seven key metrics: power consumption (W), floating-point operations per second (FLOPs), model parameters, frames per second (FPS), mean Pixel Accuracy (mPA), mean Intersection over Union (mIoU), and our novel Image Segmentation Model Benchmark (ISMB) Score. The ISMB Score serves as a standardized metric that integrates accuracy and efficiency, providing a robust measure for selecting optimal lightweight segmentation models. Our results offer valuable insights into the trade-offs between computational efficiency and segmentation performance, guiding the development of models suitable for real-time deployment in resource-constrained environments.