Impact of Localization Errors on Label Quality for Online HD Map Construction

This paper investigates how various localization errors degrade label quality in online HD map construction, revealing that heading angle errors have a more significant impact than position errors and that model performance decreases non-linearly with increasing noise, while also proposing a distance-based metric to better evaluate these effects.

The Gist

Imagine you are trying to teach a robot car how to drive itself. To do this, you need to show it a perfect, high-definition map of the world, like a super-accurate GPS that knows exactly where every lane line, curb, and crosswalk is.

The problem is, making these perfect maps is incredibly expensive and slow. It's like hiring a team of surveyors with laser equipment to walk every single street in a city.

The Big Idea: "Crowdsourcing" the Map
Instead of hiring surveyors, the researchers asked: What if we just use the data from millions of regular cars already driving around? If we take the camera and sensor data from a fleet of Uber or Tesla cars, we could theoretically build these maps for free and in real-time.

The Catch: The "Drunk Driver" Problem
Here's the issue: Regular cars don't have the super-precise GPS that surveyors have. Their location data is a bit "drunk." Sometimes they think they are 2 meters to the left of where they actually are; sometimes they think they are facing the wrong way.

If you use this "drunk" data to label the map, you are teaching the robot car with a distorted, wobbly map. The paper asks: How bad does the car's "drunk" sense of direction have to get before the robot car stops learning how to drive?

The Experiment: Three Ways to Get "Drunk"

The researchers simulated three different types of "drunk" driving errors to see how they mess up the map:

1. The "Ramp" Error (The Sudden Jump): Imagine a car driving through a tunnel (where GPS dies). It starts drifting off course slowly. When it exits the tunnel and gets GPS back, it suddenly "snaps" back to the correct position.
   • Analogy: It's like walking in a dark room, stumbling a bit, and then suddenly realizing you're standing in the wrong corner and jumping back to where you should be.
2. The "Gaussian" Error (The Jitter): This is like a shaky hand. The car's location is constantly buzzing around the true spot, like a bee hovering near a flower. It's noisy but stays generally in the right area.
   • Analogy: Trying to draw a straight line while your hand is vibrating.
3. The "Perlin" Error (The Wavy Road): This is a smooth, rolling wave of error. The car thinks it's driving in a gentle S-curve when it's actually driving straight. This happens when the car's internal math (filters) are slightly tuned wrong.
   • Analogy: Looking at a straight road through a funhouse mirror that makes everything look wavy.

The Discovery: Angles Matter More Than Distance

The researchers found a surprising truth: It doesn't matter if the car is slightly off to the left or right; it matters a lot if the car is facing the wrong way.

• Translation Error (Position): If the car thinks it's 1 meter to the left, the map is just shifted 1 meter. The robot car can usually figure this out.
• Rotation Error (Heading/Angle): If the car thinks it's facing 5 degrees to the right, the error gets massive the further away you look.
  • Analogy: Imagine you are holding a flashlight. If you move your hand 1 inch to the left, the light moves 1 inch. But if you tilt your wrist just a tiny bit, the beam of light moves miles away at the end of a long hallway.
  • Conclusion: A tiny mistake in the car's heading (which way it's pointing) creates huge, distorted errors on the map far away from the car. This is the most dangerous type of error.

The "Distance-Aware" Scorecard

Standard tests for these maps usually treat a mistake near the car the same as a mistake 50 meters away. The researchers realized this is unfair.

• The Metaphor: If you are driving, you need to know exactly where the curb is right in front of your bumper. You don't need to know the exact curve of the road 100 meters ahead right this second; you can figure that out as you get closer.
• The Solution: They created a new "scorecard" that weighs mistakes near the car much heavier than mistakes far away. This gave them a more realistic view of how well the car would actually drive.

The Results: A Little Noise is Okay, A Lot is Bad

• The "Mix" Effect: They found that if you have a dataset where 50% of the maps are perfect and 50% are "drunk," the robot car learns almost as well as if all the maps were perfect. The good data "carries" the bad data.
• The Tipping Point: However, once the noise gets too high (especially the "wavy" Perlin noise), the map becomes a hallucination. The robot car starts seeing crosswalks where there are none and lane lines that don't exist. The structure of the road completely breaks down.

The Takeaway for the Future

To build self-driving maps using regular cars, we don't need perfect surveyors. We just need to make sure the cars know which way they are facing with high precision.

If the car's GPS is a little shaky (position error), the system can handle it. But if the car's compass is a little off (heading error), the whole map becomes a distorted nightmare. The researchers suggest that future self-driving systems need to focus heavily on fixing the "heading" error, perhaps by combining data from many cars to average out the mistakes.

Technical Summary

1. Problem Statement

High-definition (HD) maps are essential for autonomous driving but are expensive and time-consuming to create using specialized survey vehicles. A promising alternative is online HD map construction using sensor data from consumer vehicle fleets. However, unlike curated datasets (e.g., Argoverse 2, nuScenes), fleet data suffers from localization errors (inaccurate ego-vehicle poses).

When these inaccurate poses are used to project global map data onto local sensor frames to create training labels, the resulting labels become distorted. The paper addresses the critical question: How do different types of localization errors affect the quality of these generated labels, and how does this degradation impact the performance of online map construction models?

2. Methodology

A. Noise Modeling

The authors introduce three distinct types of localization noise to simulate real-world localization failures, applied to the ground truth (GT) poses of the Argoverse 2 dataset:

1. Ramp Noise: Simulates GNSS signal blockage (e.g., tunnels, urban canyons) where the vehicle pose drifts linearly over time until a new fix is acquired, causing a "jump" back to the correct position. This models drift and sudden corrections.
2. Gaussian Noise: Simulates raw, unfiltered GNSS output. It applies a truncated Gaussian distribution to position and orientation, representing random, high-frequency noise without temporal consistency.
3. Perlin Noise: Simulates stable but inaccurate GNSS signals (e.g., poor Kalman filter tuning). This creates smooth, wavelike trajectory deviations rather than sharp jumps or random noise.

Note: The authors also implement a Heading Correction mechanism. When a translational offset is applied, the vehicle's heading is adjusted to match the new direction of travel, creating a more realistic angular error scenario.

B. Model and Training

• Base Model: A variant of MapTRv2, a state-of-the-art online HD map construction model using a Detection Transformer (DETR) architecture.
• Training Strategy: The models are trained on the Argoverse 2 dataset where the training labels are artificially distorted using the noise models above. The validation and test sets remain unaltered to provide a clean benchmark.
• Experimental Variables: The study varies the magnitude of translation error ($\epsilon_L$), angular error ($\epsilon_R$), and the ratio of noisy data ($NR$) in the training set.

C. Evaluation Metric: Distance-Aware Chamfer Distance

Standard metrics like Average Precision (AP) treat all map elements equally regardless of distance. The authors argue this is insufficient because:

1. Distant objects are harder to predict precisely.
2. Distant errors are less critical for immediate driving safety than near-field errors.

They propose a Distance-Aware Metric:

• The space around the vehicle is divided into concentric rings (annuli).
• Chamfer distance is calculated separately for each ring.
• This allows for a granular analysis of how label distortion affects model performance at varying perception ranges.

3. Key Contributions

1. Realistic Noise Modeling: First to model diverse localization noise (Ramp, Gaussian, Perlin) based on real-world uncertainty patterns and apply them to map perception datasets.
2. Performance Degradation Analysis: Systematically quantifies how different noise types and magnitudes degrade the performance of MapTRv2.
3. Novel Metric: Introduces a distance-aware evaluation metric that reveals how errors propagate with distance from the ego-vehicle.
4. Heading vs. Position Analysis: Demonstrates that heading (angular) errors are significantly more detrimental to map construction than pure translational (position) errors.

4. Key Results

A. Impact of Noise Types

• Angular Errors are Critical: Models trained with heading corrections (which introduce angular errors) suffer significantly more than those with only translational errors. As distance from the vehicle increases, angular errors cause massive label distortion (the "lever effect").
• Noise Severity:
  • Low Noise: Models can compensate for slight distortions (e.g., Ramp R1, Gaussian G1) with performance close to the baseline.
  • Medium Noise: Structural logic is maintained, but warping increases. Models begin to hallucinate incorrect elements (e.g., double centerlines) at the edge of the perception range.
  • High Noise: The scene structure collapses. Models output excessive duplicate predictions and lose orientation (e.g., Perlin P3).

B. Non-Linear Degradation

• The "Bad Data" Threshold: Performance does not degrade linearly with the amount of noisy data.
• Benefit of Clean Data: Even with 50% noisy data (Ramp R1), the model performs nearly as well as the baseline. However, doubling the noise to 100% causes a disproportionate drop in performance. This suggests that clean ground truth acts as a strong regularizer, allowing the model to learn the correct structure despite noisy examples.

C. Temporal Consistency

• Ramp/Perlin (Temporal Consistency) vs. Gaussian (Inconsistent): Surprisingly, for the current MapTRv2 architecture (which lacks explicit temporal memory), the performance difference between temporally consistent noise (Ramp/Perlin) and inconsistent noise (Gaussian) was minimal.
• Implication: Models with temporal dependencies (using history) might handle consistent noise (Ramp/Perlin) better than inconsistent noise, but this requires further investigation.

D. Distance-Aware Findings

• The new metric confirmed that errors in distant map elements are more severe when heading errors are present. Pure position errors have a more localized impact.

5. Significance and Conclusion

This study provides crucial guidelines for the deployment of online HD map construction using fleet data:

• Localization Precision Requirements: High-precision heading estimation is more critical than sub-meter position accuracy for map construction, especially for elements far from the ego-vehicle.
• Data Filtering: It is not necessary to discard all data with minor localization errors. As long as a significant portion of the training data is clean, the model can learn robustly. However, high levels of error (especially angular) must be filtered out to prevent structural collapse of the map.
• System Design: Localization filters (e.g., Kalman filters) must be tuned to minimize angular drift, as this has the most severe downstream impact on map perception models.

In summary, the paper establishes that while online map construction is feasible with imperfect fleet data, the quality of the heading estimate is the limiting factor for map accuracy, and the relationship between label noise and model performance is non-linear, heavily dependent on the presence of clean ground truth.