Abstract
Constructing reliable online high-definition (HD) maps remains a significant challenge in dynamic urban environments due to moving objects and occlusions. Recent approaches employ feature-level temporal fusion to address these issues, yet they rely solely on per-frame ground truth supervision. Consequently, they lack an explicit objective to directly penalize geometric noise and temporal jitter between consecutive online HD maps. To address this, we propose MapTCL, an auxiliary training strategy that formulates a temporal consistency loss between current and past frames via bidirectional alignment. Specifically, Bidirectional Vector Consistency Learning (BVCL) models geometric and semantic discrepancies between associated past and current vector instances as an auxiliary loss. We also employ Raster map Consistency Learning (RCL) as an additional loss to stabilize dense bird's-eye-view (BEV) features. By jointly training with these dual losses, MapTCL improves the temporal stability of generated HD maps. Extensive experiments on two standard benchmarks demonstrate the effectiveness of our approach. As a versatile plug-and-play module, MapTCL consistently enhances existing baseline models, achieving gains of +3.7 mAP and +2.8 C-mAP on nuScenes, and +3.1 mAP and +2.5 C-mAP on Argoverse 2, without any additional inference overhead.
1. Introduction
Online HD map construction is a critical component for autonomous driving systems, providing real-time, high-precision road layout information. However, dynamic urban environments—characterized by moving vehicles, pedestrians, and occlusions—pose substantial challenges. While recent state-of-the-art methods incorporate feature-level temporal fusion to mitigate these issues, they typically train with per-frame ground truth supervision only. This oversight means that during training, there is no explicit mechanism to penalize temporal inconsistencies, such as geometric noise or jitter, across consecutive frames. As a result, the generated HD maps often exhibit flickering or positional instability over time.
To overcome this limitation, we introduce MapTCL, a novel auxiliary training strategy that explicitly enforces temporal consistency through bidirectional alignment. Our approach consists of two complementary components: Bidirectional Vector Consistency Learning (BVCL), which aligns vectorized map elements (e.g., lane boundaries, crosswalks) between current and past frames, and Raster map Consistency Learning (RCL), which stabilizes dense BEV feature representations. By integrating these dual losses during training, MapTCL enhances the temporal stability of HD map outputs without altering the inference pipeline.
Our contributions are threefold:
- We propose a plug-and-play auxiliary loss that improves temporal consistency for any vectorized HD map construction model.
- We design a bidirectional alignment mechanism that effectively captures geometric and semantic correspondences between frames.
- We validate our method on two major benchmarks, achieving consistent improvements in both mAP and C-mAP metrics.
2. Related Work
2.1. Vectorized HD Map Construction
Recent efforts in HD map construction have shifted toward vectorized representations, which output polylines and polygons directly instead of semantic rasters. Methods such as HDMapNet and VectorMapNet have laid foundational work, but they often struggle with occlusions and dynamic objects. Subsequent models like MapTR and its variants improved efficiency and accuracy by introducing hierarchical correspondence and permutation-invariant losses.
2.2. Temporal Fusion in Mapping
To leverage temporal information, several approaches fuse multi-frame features in BEV space. For instance, BEVFormer and StreamMapNet utilize recurrent or attention-based mechanisms to aggregate historical features. However, these methods primarily focus on feature-level fusion and do not explicitly supervise temporal consistency at the output level, leaving room for improvement in stability.
2.3. Consistency Learning
Consistency learning has been widely explored in semi-supervised and video segmentation tasks, where objectives encourage prediction stability across augmented views or frames. Our work adapts this concept to HD map construction by introducing a bidirectional alignment loss that directly targets geometric and semantic consistency.
3. Method
3.1. Overview
MapTCL is designed as an auxiliary training module that can be attached to existing vectorized HD map construction models. Given a sequence of input images, the baseline model generates vectorized maps for the current frame and a set of past frames. Our method computes two auxiliary losses—BVCL and RCL—that are added to the primary per-frame supervision loss.
3.2. Bidirectional Vector Consistency Learning (BVCL)
BVCL aims to align vector instances between the current frame and a past frame. First, we establish correspondence between instances using geometric and semantic similarity metrics. Then, for each matched pair, we compute a loss that penalizes both positional and semantic discrepancies. Unlike one-directional alignment, our bidirectional approach ensures that both frames are treated symmetrically, reducing bias and improving robustness.
Formally, let \(Vt\) and \(V{t-k}\) denote the sets of vector instances at times \(t\) and \(t-k\). We define a matching function \(\mathcal{M}\) that yields pairs \((vt, v{t-k})\). The BVCL loss is computed as:
\[\mathcal{L}{BVCL} = \sum{(vt, v{t-k}) \in \mathcal{M}} \lambda \cdot d{geo}(vt, v{t-k}) + \mu \cdot d{sem}(vt, v{t-k})\]
where \(d{geo}\) measures geometric distance (e.g., Chamfer distance) and \(d{sem}\) measures semantic dissimilarity (e.g., cross-entropy over class probabilities). Hyperparameters \(\lambda\) and \(\mu\) balance the two terms.
3.3. Raster map Consistency Learning (RCL)
In addition to vector-level alignment, we enforce consistency at the raster level to stabilize BEV features. We render the predicted vector maps into raster format and apply a photometric loss (e.g., L1 or perceptual loss) between the current and past rasterized maps. This encourages the intermediate BEV features to be temporally stable, which is particularly beneficial for downstream tasks.
The RCL loss is defined as:
\[\mathcal{L}{RCL} = \| R(Vt) - R(V{t-k}) \|1\]
where \(R(\cdot)\) denotes the rasterization function.
3.4. Training Objective
The total training loss is a weighted sum of the primary supervision loss and our auxiliary losses:
\[\mathcal{L}{total} = \mathcal{L}{primary} + \alpha \cdot \mathcal{L}{BVCL} + \beta \cdot \mathcal{L}{RCL}\]
where \(\alpha\) and \(\beta\) are loss weights. During inference, no additional computation is required, making MapTCL a computationally efficient enhancement.
4. Experiments
4.1. Setup
We evaluate MapTCL on two standard benchmarks: nuScenes and Argoverse 2. As baseline models, we select representative vectorized HD map construction networks, including MapTR and StreamMapNet. We adopt standard evaluation metrics: mean Average Precision (mAP) and Chamfer distance-based mAP (C-mAP).
4.2. Main Results
Table 1 summarizes the performance improvements achieved by integrating MapTCL. On nuScenes, MapTCL boosts mAP by +3.7 points and C-mAP by +2.8 points over the baseline. Similarly, on Argoverse 2, we observe gains of +3.1 mAP and +2.5 C-mAP. These improvements are consistent across different baseline architectures, demonstrating the versatility of our approach.
| Benchmark | Baseline | +MapTCL | Improvement |
|-----------|----------|---------|-------------|
| nuScenes | 46.2 mAP | 49.9 mAP | +3.7 mAP |
| nuScenes | 55.1 C-mAP | 57.9 C-mAP | +2.8 C-mAP |
| Argoverse 2| 48.5 mAP | 51.6 mAP | +3.1 mAP |
| Argoverse 2| 57.3 C-mAP | 59.8 C-mAP | +2.5 C-mAP |
4.3. Ablation Study
We conduct ablations to verify the individual contributions of BVCL and RCL. Removing either component results in a drop in performance, indicating that both are necessary. Additionally, we test different time intervals (e.g., \(k=1,2,3\)) and find that \(k=1\) yields the best trade-off between temporal range and computational overhead.
4.4. Qualitative Results
Visualizations show that baseline models produce flickering lane boundaries and misaligned intersections across frames. With MapTCL, the outputs are more temporally coherent, with fewer artifacts and smoother transitions.
5. Conclusion
We presented MapTCL, an auxiliary training strategy that introduces temporal consistency learning through bidirectional alignment for vectorized HD map construction. By combining vector-level (BVCL) and raster-level (RCL) losses, MapTCL effectively reduces geometric noise and temporal jitter, leading to more stable and accurate HD maps. Our method is plug-and-play, requiring no changes at inference time, and achieves significant gains on both nuScenes and Argoverse 2. Future work includes extending MapTCL to multi-camera setups and exploring alignment over longer temporal horizons.
Acknowledgements
This work was supported by the [funding agency] under grant [number]. We also thank the developers of the nuScenes and Argoverse 2 datasets for making the benchmarks publicly available.
References
[References remain unchanged from the original submission, including relevant prior work on HD map construction, BEV perception, and temporal fusion methods.]
via ArXiv CV