BEV-Denoise: Learning Intrinsic Noise for Accurate Bird's-Eye-View Semantic Segmentation

The paper introduces BEV-Denoise, a framework designed to enhance Bird's-Eye-View (BEV) semantic segmentation by estimating and removing intrinsic noise from BEV features using a UNet-based noise estimation module inspired by Denoising Diffusion Probabilistic Models (DDPM). The approach employs a sequential learning paradigm called Task Decomposition, utilizing a pre-trained BEV map autoencoder to train a view transformation encoder, and achieves improved performance across four existing models on the nuScenes dataset. This advancement is significant for practitioners as it offers a method to refine BEV feature quality, potentially leading to more accurate semantic segmentation in autonomous driving and robotics applications.