DAMap: Distance-aware MapNet for High Quality HD Map Construction

TL;DR

DAMap targets two core bottlenecks in high-quality HD map construction: misaligned task labels due to one-to-many matching and sub-optimal task features from shared representations. It introduces Distance-aware Focal Loss (DAFL) to tie localization quality into classification labels, Task Modulated Deformable Attention (TMDA) to learn task-specific sampling and attention, and a Hybrid Loss Scheme (HLS) to leverage DAFL benefits across decoder layers. Empirical results on NuScenes and Argoverse2 show consistent improvements over strong baselines across multiple schedules and backbones, including larger gains with longer training and Swin backbones. Overall, DAMap demonstrates practical significance by enhancing recall and precision of high-quality HD map elements, contributing to safer autonomous driving systems.

Abstract

Predicting High-definition (HD) map elements with high quality (high classification and localization scores) is crucial to the safety of autonomous driving vehicles. However, current methods perform poorly in high quality predictions due to inherent task misalignment. Two main factors are responsible for misalignment: 1) inappropriate task labels due to one-to-many matching queries sharing the same labels, and 2) sub-optimal task features due to task-shared sampling mechanism. In this paper, we reveal two inherent defects in current methods and develop a novel HD map construction method named DAMap to address these problems. Specifically, DAMap consists of three components: Distance-aware Focal Loss (DAFL), Hybrid Loss Scheme (HLS), and Task Modulated Deformable Attention (TMDA). The DAFL is introduced to assign appropriate classification labels for one-to-many matching samples. The TMDA is proposed to obtain discriminative task-specific features. Furthermore, the HLS is proposed to better utilize the advantages of the DAFL. We perform extensive experiments and consistently achieve performance improvement on the NuScenes and Argoverse2 benchmarks under different metrics, baselines, splits, backbones, and schedules. Code will be available at https://github.com/jpdong-xjtu/DAMap.

Figures (8)

• Figure 1: Top: The analysis about recall at 0.5m threshold (high localization). Even with the 0.6 classification threshold, the recall is only $35\%$. This indicates that the baseline performs poorly in high quality predictions. Our method improves the recall of high quality predictions at each score threshold. Bottom: A visualization example with a 0.4 classification threshold. At this threshold, the baseline model fails to predict pedestrian crossings, which is dangerous for autonomous driving vehicles.
• Figure 2: The pipeline of the proposed DAMap. Taking surrounding images as inputs, a shared backbone as the feature extractor first extracts image features; The extracted image features are then transformed to BEV space as BEV features; The transform decoder finally achieves the interaction between queries and BEV features to predict instances of map elements. These predictions are used to calculate the loss for model optimization in the training phase.
• Figure 3: A simple example of inappropriate classification labels for one-to-many matching queries. The green line represents GT, and the other colored lines represent queries.
• Figure 4: Left: Frequency of the classification scores on the True Positive with or without our method. Right: Frequency of the classification scores on the False Positive with or without our method.
• Figure 5: Example of learned attention weights for points in an instance when task-shared and task modulated.