HierDAMap: Towards Universal Domain Adaptive BEV Mapping via Hierarchical Perspective Priors

TL;DR

HierDAMap addresses unsupervised domain adaptation for BEV mapping by introducing hierarchical perspective priors across global, sparse, and instance levels. It leverages a mean-teacher framework with three guiding modules—Semantic-Guided Pseudo Supervision (SGPS), Dynamic-Aware Coherence Learning (DACL), and Cross-Domain Frustum Mixing (CDFM)—to align semantic and geometric BEV representations across domains. Perspective priors originate from vision foundation models (e.g., Segment Anything with SAN) to generate pseudo-labels and masks used in BEV learning. The optimization combines losses $Loss^s$ and $Loss^t$ under EMA updates, achieving state-of-the-art cross-domain performance on nuScenes and Argoverse across semantic HD mapping, semantic mapping, and vectorized HD mapping, with comprehensive ablations and visual analyses.

Abstract

The exploration of Bird's-Eye View (BEV) mapping technology has driven significant innovation in visual perception technology for autonomous driving. BEV mapping models need to be applied to the unlabeled real world, making the study of unsupervised domain adaptation models an essential path. However, research on unsupervised domain adaptation for BEV mapping remains limited and cannot perfectly accommodate all BEV mapping tasks. To address this gap, this paper proposes HierDAMap, a universal and holistic BEV domain adaptation framework with hierarchical perspective priors. Unlike existing research that solely focuses on image-level learning using prior knowledge, this paper explores the guiding role of perspective prior knowledge across three distinct levels: global, sparse, and instance levels. With these priors, HierDA consists of three essential components, including Semantic-Guided Pseudo Supervision (SGPS), Dynamic-Aware Coherence Learning (DACL), and Cross-Domain Frustum Mixing (CDFM). SGPS constrains the cross-domain consistency of perspective feature distribution through pseudo labels generated by vision foundation models in 2D space. To mitigate feature distribution discrepancies caused by spatial variations, DACL employs uncertainty-aware predicted depth as an intermediary to derive dynamic BEV labels from perspective pseudo-labels, thereby constraining the coarse BEV features derived from corresponding perspective features. CDFM, on the other hand, leverages perspective masks of view frustum to mix multi-view perspective images from both domains, which guides cross-domain view transformation and encoding learning through mixed BEV labels. The proposed method is verified on multiple BEV mapping tasks, such as BEV semantic segmentation, high-definition semantic, and vectorized mapping. The source code will be made publicly available at https://github.com/lynn-yu/HierDAMap.

Figures (7)

• Figure 1: Visual representation and technical framework of BEV models in different domains. (a) is the visualization of the model prediction in Singapore while the model is trained on the Boston dataset. The predicted results of the model trained on Boston are unsatisfactory due to large domain gaps. (b) shows the framework of the representative previous domain adaptation paradigm PCT pct. It only employs perspective pseudo-label supervision in the whole domain at the image coding level. (c) depicts our framework. The perspective priors are hierarchically fully exploited to promote domain adaptation at different levels of the BEV mapping model.
• Figure 2: Results of different Unsupervised Domain Adaptation (UDA) methods for BEV mapping. Our method shows superior performance across various cross-domain scenarios for BEV mapping. The state-of-the-art methods DualCross man2023dualcross and PCT pct are compared.
• Figure 3: The framework of HDGMapping. The entire framework is based on mean teacher, where the student model parameters are learned from both the source and target domains, whereas the teacher model dynamically adjusts based on the student model changes, which is controlled by parameter $\alpha$.
• Figure 4: Illustration of the BEV mapping model. Based on the LSS framework LSS, it contains five modules: Image Encoder, Perspective View Head (PV Head), Perspective View Mask Generation (PV Mask Generation), View Transformer, and Multitask Head.
• Figure 5: The example diagram of cross-domain instance mixing. (a) depicts the generation of the source instance mask and BEV view mask. (b) depicts the generation process of mixed target perspective images, which is implemented by mixing instance masks from the source. (c) corresponds to the mixed BEV labels obtained from the BEV perspective mask.