GenBEV: Generative Model With Semantic Compensation for Bird’s Eye View Segmentation

Bird’s-Eye View (BEV) semantic segmentation is a key technology for constructing high-precision maps in low-cost visual navigation systems. The main challenge lies in effectively transforming image features into BEV features while preserving rich BEV visual information. Recent works have shown that generative models hold great promise in advancing BEV segmentation. However, these methods primarily focus on producing BEV features using prior knowledge, often overlooking key challenges such as feature shift, confusion, and forgetting during the BEV feature generation process. In this paper, we propose GenBEV, a generative model with semantic compensation that formally addresses inaccuracies and confusion in BEV feature generation. GenBEV leverages the synergistic benefits of data fusion consistency and noise-reduction training to enhance the diversity and reliability of the generated information. This improvement boosts the robustness and generalization of BEV segmentation across diverse scenarios, including those involving complex objects and low-quality images. Specifically, we design an adaptive cross-feature encoder to reduce diffusion variability. During decoding, we integrate the context of BEV features with noisy features to construct semantic embeddings. We show the effectiveness of GenBEV on the nuScenes, KITTI Raw, and KITTI 3D Object datasets. GenBEV achieves segmentation scores of 29.5%, 68.8%, and 39.7%, respectively, surpassing current methods by up to 3.6%, 2.4%, and 2.7%. To the best of our knowledge, GenBEV is the first to address the problem of BEV feature falsification in generative architectures.