Bird’s-Eye View Semantic Segmentation for Autonomous Driving through the Large Kernel Attention Encoder and Bilinear-Attention Transform Module

Abstract

Building an autonomous driving system requires a detailed and unified semantic representation from multiple cameras. The bird’s eye view (BEV) has demonstrated remarkable potential as a comprehensive and unified perspective. However, most current research focuses on innovating the view transform module, ignoring whether the crucial image encoder can construct long-range feature relationships. Hence, we redesign an image encoder with a large kernel attention mechanism to encode image features. Considering the performance gains obtained by the complex view transform module are insignificant, we propose a simple and effective Bilinear-Attention Transform module to lift the dimension completely. Finally, we redesign a BEV encoder with a CNN block of a larger kernel size to reduce the distortion of BEV features away from the ego vehicle. The results on the nuScenes dataset confirm that our model outperforms other models with equivalent training settings on the segmentation task and approaches state-of-the-art performance.