RO-BEV: Towards Robust BEV Feature Enhancement for Surround-View Instance Prediction

The most existing instance prediction methods often lack practical understanding and enhancement of bird's-eye view (BEV) spatial-temporal features, particularly distant features. To alleviate the above problems, we propose a robust BEV feature enhancement network (RO-BEV). The proposed RO-BEV includes two major components: the novel adaptive linearly increasing dividing (ALID) strategy and the hierarchical cross-scale fusion (HCSF) module. Specifically, the ALID strategy can mitigate instance loss caused by insufficient features by providing a linearly increasing 3D representation to build a robust BEV feature. Then, the proposed HCSF module enhances spatial-temporal BEV features in the latent space to predict vehicle instances by sampling from feature distributions at different scales and timestamps. Experimental results on the nuScenes dataset show that our RO-BEV outperforms existing state-of-the-art methods.