HoMap: End-to-End Vectorized HD Map Construction With High-Order Modeling

The construction of an online vectorized high-definition map is critical for downstream planning and decision-making in autonomous driving. Existing methods have established solid baselines for this task by modeling map elements as a collection of points that describe their shapes. However, the complex geometric relationships between map elements have not been thoroughly investigated, resulting in significant errors when dealing with issues such as multiple pedestrian crossings. In this paper, HoMap, an end-to-end online vectorized high-definition map construction framework that effectively captures complex relationships between instances is proposed. Specifically, the concept of high-order modeling is introduced for the map construction task, which involves partitioning each instance in a high-dimensional space and computing high-order statistics to capture the correlations between map elements. Then, statistical attention is incorporated as a Bird's Eye View enhancement module to reinforce semantic and geometric information. Additionally, instance queries are used as input for the decoder, which improves query capabilities and avoids inconsistencies in instance information. HoMap achieves state-of-the-art performance on the nuScenes and Argoverse2 datasets while maintaining good mapping results in complex driving environments.