UniMapGen: A Generative Framework for Large-Scale Map Construction from Multi-modal Data

UniMapGen represents a significant advancement in large-scale map construction, a critical area for technologies such as autonomous driving and navigation systems. Traditional methods often face challenges due to the high costs and inefficiencies associated with data collection and annotation. Existing satellite-based approaches, while promising, are hindered by issues like occlusions and outdated data. UniMapGen overcomes these limitations by introducing a novel framework that utilizes discrete sequences for lane line representation and supports multi-modal inputs, including bird's-eye view (BEV), perspective view (PV), and text prompts. This flexibility allows for more accurate and smoother map vector generation compared to traditional methods. The framework's effectiveness is underscored by its state-of-the-art performance on the OpenSatMap dataset, marking a pivotal step forward in enhancing the efficiency and reliability of map construction processes.