MapGCLR: Geospatial Contrastive Learning of Representations for Online Vectorized HD Map Construction

Conference: CVPR 2026 arXiv: 2603.10688 Code: None Area: Autonomous Driving Keywords: Online HD map construction, semi-supervised learning, contrastive learning, BEV features, multi-traversal

TL;DR

MapGCLR proposes a semi-supervised training scheme based on geospatial contrastive learning: it exploits the geospatial overlap between BEV feature grids produced from multiple traversals of the same location, constructing an InfoNCE contrastive loss to enforce geographic consistency in the BEV feature space. On Argoverse 2, using only 5% labeled data, it achieves 18.9 mAP (vs. 13.3 for the fully supervised baseline), a relative improvement of 42%—roughly equivalent to doubling the amount of labeled data.

Background & Motivation

Background: Online HD map construction is a critical task in autonomous driving. Methods such as MapTR/MapTRv2/MapTracker predict vectorized map elements (lane lines, road boundaries, crosswalks, etc.) from surround-view camera inputs and have achieved strong performance. However, all such methods rely heavily on large amounts of annotated training data—precise HD map annotation is extremely costly and represents the primary scalability bottleneck.

Limitations of Prior Work: - PseudoMapTrainer: Generates Gaussian surfel grid pseudo BEV segmentation labels from sensor data, but relies on the semantic segmentation paradigm rather than vectorized prediction. - Lilja et al.: Teacher-student architecture with temporal pseudo-label fusion, also based on the segmentation paradigm. - Both approaches perform semi-supervised learning via pseudo-labels without directly exploiting the natural prior of geospatial consistency.

Core Insight: In autonomous driving datasets, the same location is often traversed multiple times (multi-traversal). BEV features observed at the same geographic location under different times, weather conditions, and lighting should be similar—this constitutes a natural self-supervised signal that requires no annotation.

Key Distinction: Unlike HRMapNet/RTMap, which store and fuse historical BEV features at inference time (increasing memory and model complexity), MapGCLR exploits multi-traversal geographic consistency only during training to improve encoder representations. At inference time, it remains a single-frame, single-pass forward pass.

Method

Overall Architecture

MapGCLR builds upon MapTRv2 and adopts a dual-branch semi-supervised training pipeline:

• Supervised branch (pink): Small labeled dataset → ResNet-50 image feature extraction → BEV lifting → MapTRv2 Transformer decoder → vectorized map element predictions → supervised loss \(\mathcal{L}_{\text{sup}}\)
• Self-supervised branch (blue + orange): Large unlabeled multi-traversal data pairs (reference + adjacent) → shared encoder extracts BEV feature grids → projection head \(h\) maps to embedding space → geospatial contrastive loss \(\mathcal{L}_{\text{GCLR}}\)

Each batch contains \(n\) labeled samples and \(2m\) unlabeled samples (\(m\) reference-adjacent pairs), totaling \(n + 2m\) samples per batch.

Key Designs

1. Geospatial Multi-Traversal Analysis (Multi-Traversal Split):

   • Function: Analyzes geographic overlap between different driving logs in the dataset and partitions them into single-traversal and multi-traversal subsets.
   • Mechanism: All vehicle poses are transformed to a global coordinate frame. For each pose, a perception bounding box is computed based on vehicle heading (lateral \(\pm x\) meters, longitudinal \(\pm y\) meters). All bounding boxes within a log are merged into a polygon. A log is classified as multi-traversal if its polygon intersects with that of at least one other log.
   • Spatial Graph Construction: Graph \(G = (V, E)\), where nodes \(v \in V\) are vehicle poses and edges \(e_{ij} \in E\) connect pose pairs whose IoU falls within \([\text{IoU}_{\min}, \text{IoU}_{\max}]\). Constraining both minimum and maximum IoU ensures overlap regions are sufficiently relevant without being nearly identical.
   • Dataset Split Strategy: All multi-traversal logs are used for self-supervised training (their annotations are ignored); single-traversal logs are further divided into a supervised subset (2.5%/5%/10%/20%) and a validation set.
   • Design Motivation: In Argoverse 2, the vast majority of logs exhibit multi-traversal overlap (histograms show most logs intersect with 2–20 other logs), providing abundant positive pairs for contrastive learning.

2. Geospatial Contrastive Learning:

   • Function: Built on the SimCLR framework, constructs contrastive pairs from geographically overlapping BEV grid cells to enforce geographically consistent BEV representations.
   • Positive/Negative Sample Definition:
     • Positive pairs: A cell \(c_a\) (anchor) in the reference BEV grid paired with cell \(c_p\) in the adjacent BEV grid corresponding to the same geographic location.
     • Negative samples: Cells \(c_n\) randomly sampled from both grids that do not share a spatial correspondence with the anchor.
   • Sampling Strategy: Anchor points are randomly sampled from the overlapping region of the reference grid; the corresponding positive sample in the adjacent grid is found via nearest-neighbor search; negative samples are randomly drawn from both grids, excluding the anchor and the positive sample.
   • Core Novelty: Unlike conventional contrastive learning that constructs positive pairs via image augmentations, MapGCLR treats observations of the same location at different times as natural augmentations, leveraging real differences in viewpoint, lighting, and dynamic objects as intrinsic data augmentation.

3. Projection Head and InfoNCE Loss:

   • Function: Maps BEV cell features \(\mathbf{f}\) through projection head \(h\) into an embedding space \(\mathbf{z} \in \mathcal{Z}\), where the contrastive loss is computed.
   • Loss Function:
   \[\mathcal{L}_{\text{GCLR}} = -\log \frac{\exp(\text{sim}(\mathbf{z}_i, \mathbf{z}_i^+) / \tau)}{\exp(\text{sim}(\mathbf{z}_i, \mathbf{z}_i^+) / \tau) + \sum_{k=1}^{K} \exp(\text{sim}(\mathbf{z}_i, \mathbf{z}_k^-) / \tau)}\]

   where \(\text{sim}(\cdot, \cdot)\) denotes cosine similarity and \(\tau\) is the temperature parameter. - Design Motivation: The projection head decouples the contrastive learning domain from the downstream task domain (standard practice in SimCLR), preventing the contrastive objective from directly interfering with task-specific structure in the feature space.

Loss & Training

The total loss is a weighted combination of the supervised and contrastive losses:

\[\mathcal{L}_{\text{semi}} = \lambda_{\text{sup}} \mathcal{L}_{\text{sup}} + \lambda_{\text{GCLR}} \mathcal{L}_{\text{GCLR}}\]

• \(\lambda_{\text{sup}}\) and \(\lambda_{\text{GCLR}}\) control the relative weighting of the two objectives.
• \(\mathcal{L}_{\text{sup}}\) strictly follows the original MapTRv2 loss (Hungarian matching + classification/regression losses) and is computed only on labeled samples.
• \(\mathcal{L}_{\text{GCLR}}\) is summed over all \(m\) unlabeled multi-traversal pairs.

Key Experimental Results

Main Results

On Argoverse 2, compared against the fully supervised MapTRv2 baseline:

Label Ratio	Method	mAP	Absolute Gain	Relative Gain
2.5%	Supervised baseline	6.5	—	—
2.5%	+ SSL (MapGCLR)	8.5	+2.0	+31%
5%	Supervised baseline	13.3	—	—
5%	+ SSL (MapGCLR)	18.9	+5.6	+42%
10%	Supervised baseline	22.0	—	—
10%	+ SSL (MapGCLR)	27.3	+5.3	+24%
20%	Supervised baseline	31.0	—	—
20%	+ SSL (MapGCLR)	34.9	+3.9	+13%
30%	Supervised baseline	36.6	—	—
40%	Supervised baseline	39.8	—	—

Ablation Study

Configuration	Key Observation	Notes
Label ratio vs. gain	Fewer labels → larger gain	Relative gain of 42% at 5%, only 13% at 20%, consistent with semi-supervised learning expectations
5% + SSL vs. 10% supervised	18.9 vs. 22.0	SSL effect ≈ doubling the label budget (5% → near 10%-level performance)
PCA visualization (qualitative)	SSL feature space is cleaner	Stronger contrast at road boundaries, clearer ego-lane separation
Baseline grid artifacts	Fixed-location anomalous feature clusters	Purely supervised BEV grids exhibit artifacts at fixed coordinates in the upper-right corner; SSL eliminates this phenomenon

Key Findings

• Greatest benefit under label scarcity: At 5% labeled data, relative improvement is 42% (13.3 → 18.9 mAP), implying significant annotation cost savings in practical deployment.
• Pedestrian crossing gains are relatively smaller: Among all categories, the pedestrian crossing class benefits least (e.g., at 5%, from 7.3 to 9.9), possibly because crosswalk appearance varies considerably across time, yielding a weaker geospatial consistency signal.
• PCA analysis reveals feature space quality: After SSL training, BEV features exhibit clearer boundary separation at road edges and eliminate the fixed-coordinate artifacts present in the purely supervised model—demonstrating that contrastive learning genuinely enforces geographic consistency in the feature space.

Highlights & Insights

• Elegant problem formulation: Reinterpreting "multiple traversals of the same location" as "natural data augmentation" for contrastive learning is more grounded than artificially constructed augmentations—real variations in lighting, weather, and dynamic objects provide authentic viewpoint diversity.
• Simple and plug-and-play: The core contribution is an auxiliary training loss and a data organization strategy that leaves inference-time architecture and computation entirely unchanged. It can be directly applied to any online map construction model built on a BEV feature grid.
• Dataset analysis as a standalone contribution: The proposed multi-traversal classification methodology and geographic overlap analysis offer a new perspective on leveraging datasets such as Argoverse 2, and can be generalized to other tasks requiring spatial consistency.

Limitations & Future Work

• Validated only on the single-frame MapTRv2 model: The method has not been tested on SOTA temporal models such as MapTracker. Temporal models already exploit inter-frame consistency, so the relationship with geospatial contrastive learning—whether redundant or complementary—warrants investigation.
• Dependence on high-precision localization: Constructing contrastive pairs requires accurate global poses to compute BEV grid geographic overlap. Datasets lacking high-quality localization (e.g., nuScenes) cannot be directly accommodated.
• Operates only on the encoder: The contrastive loss backpropagates only to the BEV encoder and does not affect the Transformer decoder. Extending the self-supervised signal to the decoder side (e.g., consistency constraints on decoder queries) may yield further improvements.
• No direct comparison with other semi-supervised methods: The absence of direct experimental comparisons with PseudoMapTrainer and Lilja et al. makes it impossible to assess whether geospatial contrastive learning genuinely outperforms pseudo-label-based approaches.
• Scale limited to Argoverse 2: Multi-traversal characteristics of larger-scale datasets (e.g., nuPlan, Waymo Open) have not been analyzed, leaving scalability unverified.

Related Work & Insights

• vs. SimCLR: SimCLR constructs positive pairs via image-level augmentations; MapGCLR uses geospatial overlap to construct BEV grid-level positive pairs—transferring contrastive learning from the image domain to the BEV spatial domain.
• vs. HRMapNet/RTMap: HRMapNet maintains a global BEV feature/raster map at inference time, increasing memory and complexity; MapGCLR exploits multi-traversal data only during training, imposing zero inference overhead.
• vs. self-supervised learning in autonomous driving: This work belongs to the same paradigm as PointContrast (3D point cloud contrastive learning) and BEVDistill (BEV feature distillation)—leveraging geometric priors to construct self-supervised signals—but MapGCLR is the first to introduce cross-trajectory geospatial consistency into online map construction.
• Broader inspiration: The paradigm of "leveraging geographic consistency across multiple traversals for self-supervision" is transferable to other spatial perception tasks (3D detection, occupancy prediction, semantic segmentation), particularly in scenarios where data collection is easy but annotation is scarce.

Rating

• Novelty: ⭐⭐⭐⭐ First work to introduce geospatial contrastive learning into online vectorized map construction; the idea is novel and intuitively motivated.
• Experimental Thoroughness: ⭐⭐⭐ Ablations are clear, but validation is limited to a single model (MapTRv2) and lacks direct comparisons with other semi-supervised methods.
• Writing Quality: ⭐⭐⭐⭐ Well-structured, with rigorous method descriptions and clear figures and tables.
• Value: ⭐⭐⭐⭐ A practical solution to the annotation bottleneck, with a paradigm generalizable to a broad range of spatial perception tasks.

Related Papers

• [CVPR 2026] EMDUL: Expanding mmWave Datasets for Human Pose Estimation with Unlabeled Data and LiDAR Datasets
• [CVPR 2026] ReScene4D: Temporally Consistent Semantic Instance Segmentation of Evolving Indoor 3D Scenes
• [CVPR 2026] O3N: Omnidirectional Open-Vocabulary Occupancy Prediction
• [CVPR 2026] Failure Modes for Deep Learning-Based Online Mapping: How to Measure and Address Them
• [AAAI 2026] PriorDrive: Enhancing Online HD Map Construction with Unified Vector Priors