GenMapping: Unleashing the Potential of Inverse Perspective Mapping for Robust Online HD Map Construction

TL;DR

GenMapping tackles the generalization problem in online HD map construction by decoupling camera parameters from model learning via Inverse Perspective Mapping (IPM). It introduces a triadic architecture with a principal branch and two auxiliary branches (dense perspective and sparse OSM priors), joined by a Triple-Enhanced Merging module, and further strengthened by Cross-View Map Learning and Bidirectional Data Augmentation. The framework demonstrates state-of-the-art results in both semantic and vectorized HD mapping on nuScenes and Argoverse, with strong cross-dataset and cross-location generalization and efficient inference. These contributions offer robust, sensor-parameter-agnostic online mapping suitable for real-world autonomous driving deployments, and point to future work on occlusion handling and domain gap mitigation.

Abstract

Online High-Definition (HD) maps have emerged as the preferred option for autonomous driving, overshadowing the counterpart offline HD maps due to flexible update capability and lower maintenance costs. However, contemporary online HD map models embed parameters of visual sensors into training, resulting in a significant decrease in generalization performance when applied to visual sensors with different parameters. Inspired by the inherent potential of Inverse Perspective Mapping (IPM), where camera parameters are decoupled from the training process, we have designed a universal map generation framework, GenMapping. The framework is established with a triadic synergy architecture, including principal and dual auxiliary branches. When faced with a coarse road image with local distortion translated via IPM, the principal branch learns robust global features under the state space models. The two auxiliary branches are a dense perspective branch and a sparse prior branch. The former exploits the correlation information between static and moving objects, whereas the latter introduces the prior knowledge of OpenStreetMap (OSM). The triple-enhanced merging module is crafted to synergistically integrate the unique spatial features from all three branches. To further improve generalization capabilities, a Cross-View Map Learning (CVML) scheme is leveraged to realize joint learning within the common space. Additionally, a Bidirectional Data Augmentation (BiDA) module is introduced to mitigate reliance on datasets concurrently. A thorough array of experimental results shows that the proposed model surpasses current state-of-the-art methods in both semantic mapping and vectorized mapping, while also maintaining a rapid inference speed. The source code will be publicly available at https://github.com/lynn-yu/GenMapping.

Figures (8)

• Figure 1: Generalization analysis of HD mapping models facing cross-dataset shift. 'N2A' denotes the validation result of the model trained on the nuScenes dataset nus evaluated on Argoverse argoverse. 'A2A' follows the same definition. (a) and (b) represent the cross-dataset performance of a state-of-the-art mapping method maptrv2. Inconsistent sensor parameters between training and validation lead to projection errors, causing inaccurate detection of the positions of map instances. (c) and (d) illustrate the cross-dataset results based on the proposed method, leveraging the advantage of decoupling the sensor parameters.
• Figure 2: Mapping accuracy and generalization performance of HD map models on public datasets. The first two rows depict semantic mapping results and the last two rows depict vectorized mapping results. The figure shows the visualization results of the model trained on nuScenes nus on the validation sets of nuScenes (N2N) and Argoverse argoverse (N2A), respectively. The proposed method adopts a triadic synergy framework established with the concept of parameter decoupling, leading to stronger generalization performance.
• Figure 3: Overview of the established GenMapping framework for robust online HD map construction. The pipeline follows a triadic synergy architecture with principal and dual auxiliary branches. The triple-enhanced merging module synchronously fuses three-way features in BEV space. Bidirectional data augmentation, including forward and backward augmentation, and the cross-view map learning module are designed to enhance mapping robustness.
• Figure 4: The details of the triple-enhanced merging module. 'Conv Block' means 'Convolutional Block', that is, $\mathrm{CB}$ in Sec. \ref{['tri']}.
• Figure 5: Visualization results for semantic mapping. The proposed method is compared against state-of-the-art semantic mapping methods including HDMapNet hdmapnet and LSS LSS. Classes of divider, pedestrian, and boundary are filled with green, red, and blue.