PivotNet: Vectorized Pivot Learning for End-to-end HD Map Construction

1. Introduction

• Rasterized maps use redundant semantic pixels, ignore geometric relationships between map grids, and require post-processing to produce vectorized maps.These limitations add memory and transmission demands, computation, time, and potential accumulated errors.Source: §1 · p. 1
• Fixed-point representations can waste points or miss map details, while autoregressive decoding can increase inference time and accumulate errors.The cited examples are MapTR [15] and VectorMapNet [18], respectively.Source: §1 · pp. 1–2
• PivotNet models map elements as ordered pivotal points in a direct set-prediction framework for compact, geometry-robust vectorization.The introduction describes the representation as corner-friendly and notes that the number of pivots can vary across map elements.Source: §1 · p. 2
• Point-to-line masks encode subordinate and geometric relations, while pivot dynamic matching models connections in pivotal-point sequences.Dynamic vectorized sequence loss supervises point positions and topology using both pivot and collinear point supervision.Source: §1 · p. 2
• PivotNet is reported to outperform state-of-the-art methods on existing benchmarks.The framework is presented as end-to-end and generates vectorized representations without post-processing.Source: §1 · p. 2

2.1. Semantic map learning

• Semantic map-learning methods extract map elements from onboard sensors, while HDMapNet [11] groups pixel-wise segmentation results in post-processing to obtain vectorized maps.The discussion also describes VPN [25], LSS [26], BEVDet [8], and DETR-style [2] approaches for transforming camera features into BEV representations.Source: §2.1 · p. 2

2.2. Vectorized HD Map Construction

• MapTR [15] uses a fixed number of points, which can lose rounded-corner and right-angle details, whereas VectorMapNet [18] uses a coarse-to-fine autoregressive approach with possible long inference time and accumulated errors.PivotNet is described as using a dynamic number of pivotal points in a set-prediction paradigm.Source: §2.2 · p. 2

2.3. Topology Modeling of Map Elements

• Recurrent and bottom-up methods can be time-consuming, hard to train, or require post-processing; PivotNet instead models point- and line-level geometry uniformly with line-aware attention masks.The discussion also notes anchor-based lane priors and parameterized polynomial and Bezier curves as alternative representations.Source: §2.3 · pp. 2–3

3.1. Problem Formulation

• Map elements are ordered point sequences split into pivotal points that preserve shape and collinear points that can be removed without significantly changing it.The model seeks the shortest pivotal subsequence whose distance from the map element is below a tolerable error.Source: §3.1 · p. 3

3.2.1 Architecture Overview.

• The architecture transforms multi-view camera features into BEV features, decodes line-aware point descriptors, and predicts compact pivot-based map representations.The four components are the camera feature extractor, BEV feature decoder, line-aware point decoder, and pivotal point predictor.Source: §3.2.1 · p. 4

3.2.2 Point-Line Relation Modeling.

• The Point-to-Line Mask module shares instance-level attention masks across a line's point queries and supervises the mask to encode subordinate and geometric point-line relations.The subordinate prior distinguishes points belonging to the same or different lines; the geometric prior encodes how ordered points form a line.Source: §3.2.2 · p. 5

3.2.3 Vectorized Pivot Learning.

• Pivot Dynamic Matching finds the lowest-cost correspondence between predicted points and ground-truth pivot sequences, using a custom algorithm with O(NT) time complexity.The method fixes endpoint correspondences and uses dynamic programming to find the matching.Source: §3.2.3 · p. 5
• Dynamic Vectorized Sequence loss supervises pivot positions, collinear points between adjacent pivots, and whether predicted points are pivotal.Collinear targets are interpolated between adjacent pivots in sequence order, while pivot classification uses binary cross-entropy.Source: §3.2.3 · p. 6

3.3. Training Loss

• Training adds auxiliary BEV segmentation supervision to the dynamic sequence and line-aware losses in the overall objective.The overall loss sums DVS, line-aware, and BEV losses, with weights on the latter two.Source: §3.3 · p. 6

4.1. Experimental Settings

• Evaluation covers nuScenes and Argoverse 2, three map-element classes, a 30m front/rear by 15m lateral range, and Chamfer-distance AP.The settings also report dataset frame rates, data splits, model backbones, and training details.Source: §4.1 · pp. 6–7

4.2. Comparisons with State-of-the-art Methods

• On nuScenes, PivotNet achieves 11.2 higher mAP than existing state-of-the-art [15] at {0.2, 0.5, 1.0}m thresholds and 7.3 higher mAP at {0.5, 1.0, 1.5}m thresholds.On Argoverse 2, PivotNet is superior to existing SOTA approaches [11] with {0.2, 0.5, 1.0}m thresholds and polygon-modeled pedestrian crossings.Source: §4.2 · p. 7

4.3. Ablation Study

• The results report 8.9% higher AP on shaped road-boundaries, while the proposed point-to-line mask module is associated with 4.9% higher mAP.The nuScenes module results use thresholds {0.2, 0.5, 1.0}m.Source: Table 3 · p. 7, §4.3 · p. 8
• 1.9% mAP improvement with BEV supervision, which constrains BEV features to capture meaningful information within the BEV range.The comparison is conducted on nuScenes at {0.2, 0.5, 1.0}m thresholds.Source: Table 4 · p. 8, §4.3 · p. 8
• Hierarchical queries share line embeddings by line and point embeddings by index, whereas point queries are independent of both.The comparison concerns query design for map-element point prediction.Source: Figure 4 · p. 8
• PivotNet performs better across the evaluated AP threshold settings, with stricter thresholds described as more practical for HD maps requiring centimeter-level information.The comparison includes {0.2, 0.5}m strict thresholds and easier settings such as 1.0m and 1.5m.Source: Table 6 · p. 8, §4.3 · p. 8
• Line-aware point decoder performance improves with more layers and saturates at six, the default setting.Table 7 evaluates decoder layer count.Source: Table 7 · p. 8, §4.3 · p. 8
• Qualitative comparisons report more accurate endpoints than segmentation-based HDMapNet, smoother shapes with fewer points than MapTR, and close ground-truth alignment even at night.The visual comparison includes sunny, cloudy, rainy, and night conditions.Source: Figure 5 · p. 9, §4.3 · p. 9

• PivotNet frames map construction as a direct point-set prediction problem.Source: §5 · p. 9
• Its unified pivot-based approach models geometric and subordinate relations within map elements.Source: §5 · p. 9
• The authors report that PivotNet reaches state-of-the-art results through the Point-to-Line Mask Module, Pivot Dynamic Matching module, and Dynamic Vectorized Sequence loss.Source: §5 · p. 9

• The supplement provides additional details that the main paper omitted because of space limitations.Source: Supplementary Material · p. 10
• It includes detailed illustrations of PivotNet’s design.Source: Supplementary Material · p. 10
• It also provides additional ablation studies and extensive qualitative visualizations.Source: Supplementary Material · p. 10

A.1. Code of the custom matching algorithm

• The input cost array stores L1 distances between ground-truth points and predicted-line points, while dp[i][j] tracks the lowest matching cost for their respective prefixes.Source: §A.1 · p. 10
• The code maintains mem_sort to store the minimum cost during traversal and returns a sequence distance and matching combination.Source: §A.1 · p. 10
• When A > B, the special-case code sums diagonal costs and returns the index sequence from 0 to B-1.Source: §A.1 · p. 10

A.2. Detailed illustrations of notations

• The optimal β* selects a T-combination that splits predicted points into pivot and collinear sequences; the illustrated case uses T = 4 and N = 10.Source: §A.2 · p. 10, Figure 6 · p. 10
• T is the ground-truth sequence length and N the maximum predicted-line point count; matched pivot predictions correspond one-to-one with the T ground-truth pivots.Source: §A.2 · p. 10

A.3. Ground-truth pivot point generation

• Ground-truth pivot points are generated with Visvalingam-Whyatt simplification, which repeatedly removes the least-important interior point when its triangle area falls below a threshold.Source: §A.3 · p. 11

B.1. Impact of BEV encoder layer number

• 36.0% mAP is achieved with a single encoder and decoder layer, and PivotNet improves with more layers.Table 8 evaluates encoder-layer settings; its gray row marks the default setting.Source: §B.1 · p. 11, Table 8 · p. 11

B.2. Impact of the maximum instance number

• When evaluating maximum instance number, PivotNet’s performance saturates at (20, 25, 15) for lane-divider, ped-crossing, and road-boundary, respectively; these settings are chosen to exceed typical ground-truth instance counts.Source: §B.2 · p. 11

B.3. Impact of the maximum pivot point number

• Too few maximum pivot points limit map-element modeling, while too many increase learning burden, making the setting a trade-off.Table 10 evaluates the maximum pivot point number.Source: §B.3 · p. 11, Table 10 · p. 11

C. More qualitative visualization results

• PivotNet produces accurate, compact representations in challenging intersections and dense roads, with visual comparisons to HDMapNet and MapTR across environmental conditions.The comparisons use reproduced HDMapNet results and MapTR's released model checkpoint; Figures 7–10 show sunny, cloudy, rainy, and nighttime conditions.Source: C. More qualitative visualization… · p. 11, Figure 7 · p. 12, Figure 8 · p. 13
• At night, predictions near the vehicle closely match ground truth, and PivotNet models line segments, curves, and combinations of arbitrary shapes.Figure 10 shows the nighttime condition.Source: C. More qualitative visualization… · p. 11, Figure 10 · p. 15