Research on BEVFormer-based Underwater Object Detection and Localization Model Trained with Migrated Dataset

Abstract

Listen

Translate article

Abstract The complexity and variability of the underwater environment have presented a significant technical challenge in the detection and localization of underwater object. In this paper, an underwater object detection and localization method based on the BEVFormer model is presented and experimentally verified for effectiveness. The BEVFormer model incorporates spatial cross-attention and temporal self-attention, leveraging information from both temporal and spatial scales to enhance robustness. The Generative Adversarial Network CycleGAN was employed to generate the underwater dataset U-nuScenes, based on the nuScenes dataset. Results show that the method presented in this paper achieved 24% mAP and 0.35 NDS on U-nuScenes under the condition of utilizing only visual information.