Efficient Semantic Segmentation Backbone Evaluation for Unmanned Surface Vehicles based on Likelihood Distribution Estimation

Abstract

Obstacle detection using semantic segmentation shows a great promise for unmanned surface vehicles(USVs) in unstable marine environments. Unlike traditional machine learning, semantic segmentation models need to define suitable backbones in advance to extract features of key pixels. However, although the variety and number of backbones are massive, choosing the best one for the developer's environment in the practical application can be a daunting task. Past researches attempt to explore the ranking of backbones in specific scenarios by retraining all mainstream backbone models, which has a certain effect on some single and unchanged land scenes, but cannot be adapted to the unstable marine environment. Therefore, this paper proposes a method to quickly evaluate the suitable backbone, by extracting the representation models of different backbones without retraining and fine-tuning, separating the super-pixels of their feature distribution maps, comparing the features of different models according to likelihood distribution,and finally providing corresponding evaluation scores to give reference for backbone selection. Experimental results show that the proposed approach can provide precise backbone evaluation scores without increasing the computational effort, which can help developers quickly and accurately select the best backbone suitable for their own environment, and further design more accurate semantic segmentation models for unmanned surface vehicles.