SAR Nets: An Evaluation of Semantic Segmentation Networks with Attention Mechanisms for Search and Rescue Scenes.

Abstract

This paper evaluates four semantic segmentation models in Search-and-Rescue (SAR) scenarios obtained from ground vehicles. Two base models are used (U-Net and PSPNet) to compare different approaches to semantic segmentation, such as skip connections between encoder and decoder stages and using a pooling pyramid module. The best base model is modified by including two attention mechanisms to analyze their performance and computational cost. We conduct a quantitative and qualitative evaluation using our SAR dataset defining eleven classes in disaster scenarios. The results demonstrate that the attention mechanisms increase model performance while minimally affecting the computation time.