Abstract
High-definition (HD) maps are crucial for autonomous vehicles, but their creation and maintenance is very costly. This motivates the idea of online HD map construction. To provide a continuous large-scale stream of training data, existing HD maps can be used as labels for onboard sensor data from consumer vehicle fleets. However, compared to current, well curated HD map perception datasets, this fleet data suffers from localization errors, resulting in distorted map labels. We introduce three kinds of localization errors, Ramp, Gaussian, and Perlin noise, to examine their influence on generated map labels. We train a variant of MapTRv2, a state-of-the-art online HD map construction model, on the Argoverse 2 dataset with various levels of localization errors and assess the degradation of model performance. Since localization errors affect distant labels more severely, but are also less significant to driving performance, we introduce a distance-based map construction metric. Our experiments reveal that localization noise affects the model performance significantly. We demonstrate that errors in heading angle exert a more substantial influence than position errors, as angle errors result in a greater distortion of labels as distance to the vehicle increases. Furthermore, we can demonstrate that the model benefits from non-distorted ground truth (GT) data and that the performance decreases more than linearly with the increase in noisy data. Our study additionally provides a qualitative evaluation of the extent to which localization errors influence the construction of HD maps.
Introduction
High-definition (HD) maps provide critical environmental context for autonomous vehicles, yet their high cost of creation and maintenance remains a significant barrier to scaling. Online HD map construction, which leverages onboard sensor data from consumer vehicle fleets to build maps dynamically, offers a promising solution. However, unlike curated datasets, fleet data is subject to localization errors from consumer-grade GPS/IMU systems. These errors distort the map labels used for training. This study investigates how different types of localization noise—Ramp, Gaussian, and Perlin—affect the performance of state-of-the-art online map construction models.
Methodology
The researchers employed MapTRv2, a transformer-based model that predicts vectorized map elements. To simulate real-world conditions, they injected synthetic noise into the training labels of the Argoverse 2 dataset.
Three noise profiles were tested:
- Ramp Noise: Simulates gradual drift followed by sudden corrections.
- Gaussian Noise: Represents high-frequency sensor jitter.
- Perlin Noise: Models smooth, wave-like oscillations typical of filtered localization outputs.
To evaluate performance, the study introduces a Distance-aware Map Construction Metric ($CD_\odot$). Unlike standard metrics, $CD_\odot$ calculates the Chamfer Distance within concentric rings around the vehicle, prioritizing nearby accuracy which is more critical for safe driving.
Results and Analysis
The experiments reveal that localization noise significantly degrades model performance, often in a non-linear fashion. The baseline model achieved a Mean Average Precision (mAP) of 45.8, which dropped to 38.9 with Ramp noise (2m translation, 1° heading) and as low as 27.0 under severe Perlin noise.
Heading vs. Translation Errors
A key finding is that heading (angular) errors are substantially more destructive than translation (positional) errors. While a translation error causes a uniform shift, a heading error causes the discrepancy between the label and the ground truth to diverge as the distance from the vehicle increases. This makes long-range perception particularly vulnerable to orientation inaccuracies.
Structural Impact
Qualitative evaluations demonstrate that high noise levels lead to "structural collapse." As shown in Figure 5, the model fails to maintain scene topology under heavy noise, often predicting overlapping centerlines or hallucinating features like pedestrian crossings.
The Value of Clean Data
Training on a hybrid dataset (50% clean, 50% noisy) yielded significantly better results than training on 100% moderately noisy data. This suggests that even a limited amount of high-precision data can serve as a geometric "anchor," helping the model learn the underlying logic of road geometry despite the surrounding noise.
Conclusion
This research quantifies the sensitivity of online HD map construction to localization quality. The findings highlight the critical need for robust heading estimation in fleet vehicles used for map data collection. Future work will explore temporal models that might mitigate high-frequency jitter and the potential for multi-drive aggregation to cancel out individual sensor drifts.
Summary updated