RelMap: Spatial Reasoning & Map Construction

• RelMap is a framework that employs spatial relationships and neural architectures for HD map construction and sensor data interpolation.
• It leverages class-aware spatial priors, transformer decoders, and Mixture-of-Experts to enhance map accuracy and semantic interpretability.
• The approach integrates GNNs, adaptive sensor densification, and uncertainty quantification for robust spatiotemporal visualization.

RelMap designates multiple distinct frameworks in machine learning and computer vision for map reasoning, spatiotemporal sensor field interpolation, and online high-definition (HD) map construction. The common theme among these approaches is the explicit modeling or utilization of spatial relationships—whether among sensors, map elements, or vectorized entities—and the integration of advanced neural network architectures to enhance data fidelity, imputation, and downstream interpretability. Two notable instantiations are: (1) RelMap for online HD map construction via class-aware spatial and semantic priors (Cai et al., 29 Jul 2025), and (2) RelMap for reliable spatiotemporal sensor data visualization through imputative spatial interpolation and uncertainty-aware heatmaps (Chen et al., 2 Aug 2025).

1. Architectural Foundations and Problem Domains

The RelMap framework for HD map construction (Cai et al., 29 Jul 2025) builds atop query-based, deformable-Transformer decoders, extending prior lines such as MapQR (single-frame) and MapTracker (temporal). Its core inputs are multi-view images; intermediate processing employs ResNet-50 backbones followed by BEV (bird’s-eye view) encoders, culminating in a transformer-based decoder enhanced with spatial relation and semantic expert priors. Outputs are vectorized map instances, each delineated as an ordered set of 2D points.

In spatiotemporal sensor data visualization (Chen et al., 2 Aug 2025), RelMap targets the conversion of irregular, sparsely sampled sensor networks into continuous, uncertainty-aware spatial heatmaps. The method densifies the field with virtual sensors, imputes their readings using a domain-informed GNN, and interpolates onto a regular grid via radial basis functions (RBFs), while rigorously quantifying uncertainty from placement, reliability, and model error.

2. Class-Aware Spatial Relation Priors and Semantic Expert Modulation

In (Cai et al., 29 Jul 2025), decoder self-attention is augmented by a learnable relation bias $R^{\mathrm{rel}} \in \mathbb{R}^{N_{\mathrm{ins}} \times N_{\mathrm{ins}} \times N_{\mathrm{head}}$, capturing both geometric and semantic dependencies. Each predicted instance $i$ is represented by its axis-aligned bounding box $b_i = (cx_i, cy_i, w_i, h_i)$; for each pair, log-relative normalized features $\delta x, \delta y, \delta w, \delta h$ are computed and sinusoidally encoded, yielding $R^{\mathrm{spatial}}_{i,j}$. Class-aware modulation introduces a learnable tensor $\mathbf R_{cls}$; pairwise class-probability embeddings are aggregated via the soft class-probabilities from the previous decoder layer, contributing a class-conditioned term $R^{\mathrm{cls}}_{i,j}$. These two biases sum to form $R^{\mathrm{rel}}_{i,j}$, directly modifying each attention score as:

$$\mathrm{Attn}_{\mathrm{self}}(\mathbf Q, \mathbf K, \mathbf V) = \mathrm{softmax}\left(\frac{\mathbf Q \mathbf K^\top}{\sqrt{d}} + R^{\mathrm{rel}}\right)\mathbf V$$

The final decoder feed-forward block employs a Mixture-of-Experts (MoE) prior: each instance’s feature $\mathbf x_i$ is routed to $i$0 class-specific experts $i$1, aggregated as $i$2, where $i$3 is the instance’s class probability. This mechanism enables class-adaptive feature refinement without explicit routing losses or additional routers.

3. Spatiotemporal Sensor Imputation via GNNs

For spatiotemporal sensor interpolation (Chen et al., 2 Aug 2025), the sensor field is modeled as a fully connected graph $i$4 at each time window, with edge weights $i$5 based on the Haversine distance among the $i$6 nearest neighbors. The first GNN layer restricts message passing to original sensors; subsequent layers consider all (physical and virtual) nodes. Node features are $i$7-length time series. The GNN alternates spatial Principal Neighborhood Aggregation (PNA) convolutions—which blend $i$8 distance-invariant (mean, softmax-pool, softmin-pool, standard deviation) and distance-variant (weighted mean/variance) functions—and temporal 1D convolutions. Geographic awareness is enforced through Geographical Positional Encoding (GPE): latitude/longitude are rescaled, projected to a Fourier basis over $i$9 scales, and linearly mapped, ensuring the GNN aligns with spatial structure.

The network is trained under random sensor masking using the Huber loss, optimizing imputation accuracy over masked entries while balancing sensitivity and robustness.

4. Adaptive Sensor Densification and RBF Interpolation

Crucial to reliable field interpolation is adaptive densification. A 2D kernel density estimate $b_i = (cx_i, cy_i, w_i, h_i)$0 is inverted via $b_i = (cx_i, cy_i, w_i, h_i)$1 to oversample low-density regions. Candidate virtual sensors (in number $b_i = (cx_i, cy_i, w_i, h_i)$2) are placed and iteratively adjusted through Centroidal Voronoi Tessellation to minimize clustering and ensure spatial coverage.

Imputed values for all (original/virtual) sensors yield $b_i = (cx_i, cy_i, w_i, h_i)$3 reference sites; RBF interpolation then constructs the continuous field:

$b_i = (cx_i, cy_i, w_i, h_i)$4

where $b_i = (cx_i, cy_i, w_i, h_i)$5 and the linear system is regularized by a smoothing term $b_i = (cx_i, cy_i, w_i, h_i)$6. Parameters are fixed at $b_i = (cx_i, cy_i, w_i, h_i)$7, $b_i = (cx_i, cy_i, w_i, h_i)$8, and $b_i = (cx_i, cy_i, w_i, h_i)$9 nearest neighbors per location in all experiments.

5. Uncertainty Quantification and Visualization

RelMap (Chen et al., 2 Aug 2025) explicitly parses uncertainty stemming from (S₁) sensor placement, (S₂) per-sensor reliability, and (S₃) model-interpolation error. Visualization employs extrinsic, static glyphs:

• Diagonal hatching: local opacity inversely proportional to $\delta x, \delta y, \delta w, \delta h$0, highlighting low-coverage uncertainty (S₁).
• Coarse-grid “reliability glyphs”: in each cell $\delta x, \delta y, \delta w, \delta h$1, the primary arrow height $\delta x, \delta y, \delta w, \delta h$2 indicates mean imputation error, auxiliary arrowheads mark $\delta x, \delta y, \delta w, \delta h$3/$\delta x, \delta y, \delta w, \delta h$4 percentiles (S₂), and arrow width $\delta x, \delta y, \delta w, \delta h$5 signals spatial extrapolation distance (S₃). Glyphs remain grayscale for composability atop arbitrary heatmaps and are non-interactive.

6. Empirical Performance and Ablations

In online HD map construction (Cai et al., 29 Jul 2025), RelMap-SF achieves 68.4% mAP at 24 epochs and 74.0% at 110 epochs on nuScenes, outperforming MapTRv2 by +6.9 mAP. Temporal RelMap-TF (72 epochs) attains 77.1% (vs. MapTracker 76.1%, MapExpert 76.5%). On Argoverse 2, RelMap-SF yields 69.9% (2D) and 68.5% (3D) mAP, outperforming MapTRv2 and matching or slightly exceeding HIMap. Ablation studies show incremental improvements from spatial encoding (+0.6), class-aware modulation (+0.9), and MoE (+1.3); the class-probability–driven MoE design is optimal compared to vanilla or top-2 routed alternatives.

For spatiotemporal imputation (Chen et al., 2 Aug 2025), RelMap’s GNN minimizes RMSE/MAE against IGNNK, Ordinary Kriging, and KNN on 10/12 evaluated cases, achieving up to 50% relative improvement. The reference-enhanced RBF interpolation increases SSIM by 5–15% (over direct RBF), particularly in zones with abrupt spatial events. Joint imputation plus super-resolution (factors up to $\delta x, \delta y, \delta w, \delta h$6) reduces RMSE/MAE by 10–90% relative to piecewise-linear temporal interpolation. GNN components (PNA or GPE) each degrade model performance by 10–30% if removed. User studies demonstrate high interpretability and subjective usefulness (Likert median 4.33–4.57 for understandability/usefulness).

7. Limitations and Future Research

Current RelMap variants for HD map construction assume univariate, scalar field data and do not address vector outputs (e.g., wind, currents) or multivariate sensor fusion (e.g., inclusion of remote sensing or land-cover covariates). Class-aware relation and MoE priors operate without external or hand-crafted priors, but richer statistical uncertainty models (e.g., posterior variances) remain unexplored. For online map construction, further accuracy gains may arise from more structured MoE designs or deeper exploitation of vector topologies. In spatiotemporal imputation, extension to richer statistical models or interactive, dynamic glyphs may enable even greater decision support fidelity.

---

In summary, RelMap frameworks exemplify state-of-the-art approaches for both online map construction and uncertainty-quantified spatiotemporal field visualization by systematically integrating relational priors, expert semantic modulation, advanced GNN architectures, and adaptive spatial reasoning. These contributions significantly advance spatial AI’s robustness and interpretability under realistic, resource-constrained, and sparsely sensed conditions (Cai et al., 29 Jul 2025, Chen et al., 2 Aug 2025).