DAMap: Quality Alignment in HD Map Construction

• DAMap is a method for high-quality HD map construction in autonomous driving that integrates Distance-aware Focal Loss, Hybrid Loss Scheme, and Task Modulated Deformable Attention to align classification and localization.
• It mitigates inherent misalignment defects by employing soft labeling and decoupled attention mechanisms, enhancing both semantic evidence and geometric accuracy.
• Evaluations on nuScenes and Argoverse2 demonstrate that DAMap significantly improves strict localization thresholds and overall map element prediction quality.

Searching arXiv for the primary DAMap paper and closely related naming variants to ground the article. DAMap most commonly denotes “Distance-aware MapNet for High Quality HD Map Construction,” an online HD map construction method from multi-camera inputs that targets the misalignment between classification confidence and localization quality in transformer-based vectorized mapping systems (Dong et al., 26 Oct 2025). In that usage, DAMap attributes degraded high-quality prediction to two inherent defects: inappropriate task labels under one-to-many matching, and sub-optimal task features caused by task-shared sampling. The method addresses these defects through Distance-aware Focal Loss (DAFL), Hybrid Loss Scheme (HLS), and Task Modulated Deformable Attention (TMDA). The same string, or closely related variants, has also been used in other contexts, including distribution-guided Mapper methods in topological data analysis, occupancy mapping, anti-disturbance HD map reconstruction, and cooperative perception, so the term is context-dependent rather than globally unique (Tao et al., 2024).

1. Terminology and scope

In autonomous driving literature, DAMap refers to “Distance-aware MapNet for High Quality HD Map Construction,” a method developed to improve both classification and localization quality for vectorized HD map prediction (Dong et al., 26 Oct 2025). Its stated objective is high-quality prediction of HD map elements, where “high quality” means high classification and localization scores. The method is formulated for modern BEV-based, DETR-style vectorized map construction pipelines and is evaluated on NuScenes and Argoverse2.

The same or similar naming appears elsewhere with different meanings. In topological data analysis, “DAMap” is used as a naming variant for a distribution-guided Mapper algorithm also referred to as D-Mapper, which replaces fixed interval lengths and fixed overlap ratios with density-guided covers derived from a mixture model on the lens values (Tao et al., 2024). A related density-sensitive Mapper framework, Multimapper, implements local scale refinement and persistence-based diagnostics without defining a single explicit density estimator (Deb et al., 2019). In robotics, D-Map denotes an occupancy grid mapping framework without ray-casting for high-resolution LiDAR sensors (Cai et al., 2023). In online vectorized HD map reconstruction, ADMap denotes an anti-disturbance framework with Multi-Scale Perception Neck, Instance Interactive Attention, and Vector Direction Difference Loss (Hu et al., 2024). AdaMap, by contrast, is a cooperative perception system for connected and automated vehicles (Liu et al., 2023). Accordingly, DAMap is best treated as a disambiguated term whose precise meaning must be inferred from domain and citation context.

2. Problem setting and misalignment diagnosis

DAMap is motivated by a specific failure mode in transformer-based vectorized mapping systems: many predictions with high classification scores have poor geometry, which lowers recall at strict distance thresholds (Dong et al., 26 Oct 2025). The paper locates the source of this behavior in “inherent task misalignment,” and identifies two causes.

The first cause is inappropriate task labels under one-to-many matching. In MapTRv2-style training, each ground-truth element can supervise multiple matched queries, for example seven, but all such matched queries receive the same positive class label even when their localizations differ substantially. The consequence is that low-quality, poorly localized queries are still encouraged to predict high confidence. The second cause is sub-optimal task features due to task-shared sampling. In the decoder, the same deformable cross-attention sampling mechanism is used for both classification and localization, although the two tasks prefer different evidence: semantics for classification and geometry for localization.

The pipeline to which DAMap is applied is standard for contemporary online HD map construction from images. Synchronized multi-view RGB images are processed by a shared 2D backbone, such as ResNet-50 by default or Swin-B in stronger configurations, then transformed into BEV features with a pre-defined grid size $H \times W$, producing

$F \in \mathbb{R}^{C \times H \times W}.$

A DETR-like transformer decoder with self-attention, deformable cross-attention, and feed-forward layers interacts learned queries with BEV features to produce per-element instance features and predictions. Map elements are represented as vectorized polylines, including road boundary, lane divider, pedestrian crossing, and in additional experiments lane centerline. DAMap does not replace this overall paradigm; rather, it intervenes at the loss design and decoder cross-attention level to improve alignment between scores and geometry.

3. Architectural components

DAMap consists of three named components: DAFL, HLS, and TMDA (Dong et al., 26 Oct 2025). Their roles are complementary rather than redundant. DAFL modifies classification supervision, TMDA modifies feature extraction within the decoder, and HLS determines where DAFL should be applied across decoder depth.

Component	Function	Inference impact
DAFL	Uses a soft positive label reflecting each matched query’s localization quality	Training-only
HLS	Uses standard FL in early decoder layers and DAFL in later layers	No inference cost
TMDA	Decouples cross-attention feature sampling for classification vs. localization	Active in training and inference

DAFL is introduced to align classification targets with localization quality. Instead of assigning the same binary positive label to all matched positives, it computes a distance-based localization confidence for each matched query and uses that value as a soft label in $[0,1]$. This changes classification from purely categorical supervision into quality-aware supervision.

TMDA modifies deformable cross-attention. Rather than using a single task-shared attention mechanism, it splits the query representation into classification-aware and localization-aware parts after self-attention, then applies task-specific attention weights while keeping offsets shared. The design rationale given in the paper is that offsets are harder to optimize and unbounded, so sharing them improves stability, whereas task-specific weights allow the model to attend differently for semantic and geometric evidence.

HLS is a decoder-depth scheduling rule. The paper reports that directly replacing focal loss with DAFL in all decoder layers is less effective because early-layer queries are poorly localized and randomly initialized. HLS therefore keeps standard focal loss in early layers and applies DAFL only in later layers, exploiting the fact that later decoder layers have better geometry.

The decoder formulation is summarized in the paper as

$$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$

with normalization omitted in the shorthand. Classification and localization heads operate on the decoded instance features as

$$\mathbf{O}_{\mathrm{cls}} = \mathrm{H}_{\mathrm{cls}}(\hat{\mathbf{Q}}), \qquad \mathbf{O}_{\mathrm{loc}} = \mathrm{H}_{\mathrm{loc}}(\hat{\mathbf{Q}}).$$

This placement makes DAMap an intervention inside a familiar transformer decoder stack rather than a separate map-construction pipeline.

4. Loss design, matching, and task-modulated attention

The technical core of DAMap lies in how it converts geometric error into classification supervision and how it separates task-specific attention patterns (Dong et al., 26 Oct 2025).

The baseline focal loss is

$$\text{FL}(p,y)= \begin{cases} -(1-p)^\gamma \log(p), & y=1 \ -p^\gamma \log(1-p), & y=0, \end{cases}$$

where $p$ is the predicted foreground probability and $y \in \{0,1\}$.

DAMap first defines an instance localization loss

$\mathcal{L}_{\mathrm{dist}} = \sum_{i=0}^{N-1} \mathbbm{1}_{\{c_i \neq \varnothing\}} \sum_{j=0}^{N_v-1} D_{\mathrm{L1}}\big(\hat{v}_{\hat{\pi}(i),j},\, v_{i,\hat{\gamma}_i(j)}\big).$

Here, $\hat{\pi}$ is the instance-level assignment, $F \in \mathbb{R}^{C \times H \times W}.$0 is the point-level matching for instance $F \in \mathbb{R}^{C \times H \times W}.$1, and $F \in \mathbb{R}^{C \times H \times W}.$2 is the L1 distance between matched points. DAMap then maps that distance-based loss to a localization confidence,

$F \in \mathbb{R}^{C \times H \times W}.$3

with $F \in \mathbb{R}^{C \times H \times W}.$4 controlling the rate of decay. A perfectly localized instance receives confidence $F \in \mathbb{R}^{C \times H \times W}.$5, and confidence decreases as localization error grows.

Distance-aware Focal Loss replaces binary positive labels with these soft labels:

$F \in \mathbb{R}^{C \times H \times W}.$6

For negatives, $F \in \mathbb{R}^{C \times H \times W}.$7, so the background behavior is preserved. For positives, the label becomes quality-aware. Under one-to-many matching, this means different matched queries for the same ground-truth element receive different positive targets, which removes the ambiguity produced by assigning them all the same label.

HLS defines classification supervision across decoder layers. The three forms listed in the paper are:

$F \in \mathbb{R}^{C \times H \times W}.$8

$F \in \mathbb{R}^{C \times H \times W}.$9

and the hybrid form

$[0,1]$0

with $[0,1]$1. The reported best configuration uses focal loss in only the first decoder layer and DAFL afterward.

TMDA begins by doubling the channel dimension of the query embedding and splitting it after self-attention:

$[0,1]$2

Standard deformable attention is written as

$[0,1]$3

$[0,1]$4

DAMap modifies this into task-aware attention:

$[0,1]$5

$[0,1]$6

$[0,1]$7

This preserves shared offsets but makes attention weights task-specific.

5. Empirical performance and ablation results

DAMap is evaluated on nuScenes and Argoverse2 under both “easy” and “hard” Chamfer Distance threshold sets, and the reported results show consistent improvement across backbones, schedules, and baselines (Dong et al., 26 Oct 2025). The easy thresholds are $[0,1]$8 m, and the hard thresholds are $[0,1]$9 m. A detection is treated as a true positive if its Chamfer Distance is below the threshold.

Setting	Baseline	With DAMap
nuScenes, ResNet-50, 24 epochs, MapTRv2 repro	60.4 easy / 36.6 hard	62.8 easy / 39.0 hard
nuScenes, ResNet-50, 24 epochs, MapQR repro	66.4 easy / 43.3 hard	68.8 easy / 46.0 hard
nuScenes, ResNet-50, 110 epochs, MapTRv2 repro	68.3 easy	70.4 easy / 47.4 hard
Argoverse2, ResNet-50, 6 epochs, MapTRv2 repro	63.6 easy / 38.1 hard	66.2 easy / 40.9 hard
Argoverse2, ResNet-50, 24 epochs, MapTRv2 repro	68.5 easy / 43.5 hard	69.4 easy / 45.0 hard

The gains are especially notable on the hard thresholds, which the paper interprets as evidence of better localization quality and better alignment between classification and geometry. On a new nuScenes split following the StreamMapNet protocol, MapTRv2 increases from 26.9 mAP to 28.6 mAP with DAMap, and StreamMapNet increases from 34.1 mAP to 35.6 mAP. When the lane centerline category is added, MapTRv2 improves from 53.2 mAP to 58.3 mAP on nuScenes and from 58.5 mAP to 62.5 mAP on Argoverse2.

Ablation results isolate the contribution of each component. On nuScenes with ResNet-50 for 24 epochs, the MapTRv2 reproduction baseline is 60.4 easy mAP. Adding DAFL in all layers yields 60.9, HLS yields 61.6, TMDA alone yields 61.6, and the full DAMap configuration reaches 62.8. This supports the paper’s central claim that alignment requires both loss-level and feature-level changes. HLS is further justified by a decoder-depth study showing that using focal loss only in the first layer and DAFL in the remaining layers gives better performance than using DAFL everywhere.

TMDA design ablations compare several variants with the same 52M parameter count. The reported best version is the one adopted in DAMap: shared offsets computed from concatenated task queries, with task-specific attention weights. It reaches 61.6 mAP, compared with 60.7, 60.9, and 60.4 for alternative settings. The paper therefore argues that decoupling weights, rather than offsets, is the most effective use of the added task structure.

Implementation details place DAMap close to existing MapTRv2 settings. The reported training configuration uses batch size 16, learning rate $$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$0 with cosine decay, and 4 $$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$1 RTX 3090 GPUs. TMDA increases parameters from 40M to 52M and slightly reduces speed, for example from 12.6 FPS to 12.1 FPS in the 110-epoch ResNet-50 setting, while HLS and DAFL add no inference cost.

6. Related usages and common confusions

The most common misconception surrounding DAMap is that it denotes a single method across the literature. In fact, the name collides with several independent research threads.

In topological data analysis, D-Mapper introduces a distribution-guided Mapper algorithm that fits a mixture model to scalar lens values, constructs intervals by component quantiles

$$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$2

and evaluates outputs with an adjusted silhouette coefficient

$$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$3

with $$\hat{\mathbf{Q}} = \{\mathrm{SA}(\mathbf{Q}), \mathrm{CA}(\mathbf{Q}, \mathbf{F}), \mathrm{FFN}(\mathbf{Q})\},$$4 by default (Tao et al., 2024). Multimapper, in turn, provides density-sensitive local refinement and persistence-based diagnostics through piecewise covers and nerve gluing, and can be viewed as a practical density-aware Mapper framework without committing to a single explicit density estimator (Deb et al., 2019).

In robotics, D-Map is an occupancy grid mapping framework for high-resolution LiDAR sensors that avoids ray-casting by using a depth image, an efficient on-tree update strategy, and removal of known cells to obtain a decremental property (Cai et al., 2023). Its focus is occupancy inference, segment-tree queries, and octree traversal rather than vectorized HD map element prediction. In autonomous driving, ADMap is a different online vectorized HD map reconstruction framework designed to mitigate point-order jitter through Multi-Scale Perception Neck, Instance Interactive Attention, and Vector Direction Difference Loss (Hu et al., 2024). AdaMap is again distinct: it is a real-time cooperative perception system for connected and automated vehicles with tightly coupled data and control planes (Liu et al., 2023).

These distinctions matter because the shared naming can obscure substantive differences in problem formulation. The autonomous-driving DAMap of (Dong et al., 26 Oct 2025) is specifically a quality-alignment method for vectorized HD map construction; the TDA usage concerns Mapper covers and topological summaries; D-Map concerns occupancy mapping without ray-casting; ADMap concerns anti-disturbance point-sequence stabilization; and AdaMap concerns cooperative perception scalability. The overlap is nominal rather than methodological.