NeoMap: Map-Centric Neural Representations

• NeoMap is a family of map-centered neural representations that integrate latent mapping, semantic priors, and geometric data across various applications.
• It employs structured priors, graph correspondences, and manifold constraints to fuse sparse observations and continuously update mapping information.
• NeoMap frameworks include training-free novel-view synthesis techniques that leverage pre-trained video models for photorealistic, view-consistent outputs.

Searching arXiv for papers on “NeoMap” and closely related usages to ground the article. Tool unavailable in this environment, so I will ground the article strictly in the supplied arXiv records and cite them directly. “NeoMap” is used in recent arXiv literature to denote, or to frame, several technically distinct systems concerned with maps as latent representations, semantic priors, correspondence structures, or generative constraints. One usage names a training-free framework for monocular novel-view synthesis from pre-trained video models (Li et al., 2 Jul 2026). Other usages describe “NeoMap”-style systems for global 2D semantic map construction from sparse street imagery, neural HD map priors for autonomous driving, neural map growing in Bird’s-Eye View (BEV), map-to-map matching, landmark spectral embeddings, neural surface maps, and time-aware encyclopedic globe interfaces (Toso et al., 2024, Xiong et al., 2023, Zhu et al., 2023, Ying et al., 23 Aug 2025, Sule et al., 10 Feb 2025, Morreale et al., 2021, Oliva, 2012). This suggests that the term functions less as a single standardized object than as a recurring label for map-centric neural representations and alignment mechanisms across computer vision, geometry processing, and spatiotemporal information systems.

1. NeoMap as a family of map-centered neural representations

In autonomous driving and visual mapping, “NeoMap” is repeatedly associated with replacing or augmenting explicit geometric maps by compact learned representations. “Maps from Motion” characterizes its own method as a “semantic SfM in 2D”: it reconstructs a global 2D semantic object map from sparse, uncalibrated multi-view images, without doing full 3D SfM or explicit feature matching (Toso et al., 2024). “Neural Map Prior for Autonomous Driving” defines a global neural map prior as a learned feature grid over the city that stores latent BEV feature vectors for each location, can be queried given the ego’s global pose, and can be updated by fusing new observations from new trips (Xiong et al., 2023). “NeMO: Neural Map Growing System for Spatiotemporal Fusion in Bird’s-Eye-View” describes a readable and writable big map whose grid cells are updated over time by a shared neural fusion mechanism (Zhu et al., 2023).

A related but distinct use appears in geometry processing. “Neural Surface Maps” advocates considering neural networks as encoding surface maps, with each surface patch represented by a neural UV-to-3D map and inter-surface maps represented by additional neural networks over a canonical planar domain (Morreale et al., 2021). In spectral graph learning, “Neumann eigenmaps for landmark embedding” presents NeuMaps, a landmark-based variant of diffusion maps that builds a Neumann Laplacian on a landmark subgraph and yields an embedding with a built-in Nyström extension (Sule et al., 10 Feb 2025). In graph matching, “UM3: Unsupervised Map to Map Matching” proposes an unsupervised graph-based framework for aligning heterogeneous road-network maps under sparse features, noisy coordinates, and large-scale processing constraints (Ying et al., 23 Aug 2025).

A broader historical precursor is “Project G.N.O.S.I.S.”, which envisions a time-aware, multi-layer encyclopedic globe in which the Earth is “dress[ed] up” with time-varying “skin-maps” for geology, genetics, agriculture, ethnology, linguistics, musicology, metallurgy, and other domains (Oliva, 2012). The commonality across these lines of work is not a shared algorithmic core, but an emphasis on maps as computational state: persistent, queryable, updatable, and often fused from incomplete observations.

2. Semantic and HD mapping in street-level and autonomous-driving settings

A central “NeoMap”-like interpretation in the supplied literature is automatic semantic map construction from images and vehicle traversals. In “Maps from Motion”, the input is a sparse set of street-level images, unordered, uncalibrated except for nominal intrinsics, and without reliable GPS or pose; the output is a global 2D semantic map of static urban objects with 2D ground-plane coordinates and estimated 2D camera locations (Toso et al., 2024). The method converts each image into a local top-down semantic map centered on the camera, then aligns these local maps through a graph representation whose nodes are detections and cameras, whose intra-image edges preserve topology, and whose inter-image same-class edges represent all possible correspondences. The graph encodes local 2D coordinates, one-hot semantic labels, and bounding-box features, and a GNN predicts global 2D coordinates for all nodes jointly. On real-world data, the best models reach global 2D registration with average accuracy within 4 meters even on sparse sequences with strong viewpoint change, while COLMAP has an 80% failure rate on the small sparse regime (Toso et al., 2024).

In HD map learning, “Neural Map Prior for Autonomous Driving” defines a global prior

$p^G \in \mathbb{R}^{H_G \times W_G \times C},$

a latent BEV feature map over the city rather than an explicit semantic raster (Xiong et al., 2023). Local BEV features from the current sensor input are fused with features sampled from this prior using current-to-prior cross-attention, where current BEV features act as queries and prior BEV features act as keys and values. The prior is then updated through a 2D convolutional GRU: $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$ This framework is reported to be compatible with HDMapNet, LSS, BEVFormer, and VectorMapNet, with gains of $+4.32$, $+5.02$, and $+5.53$ mIoU for raster segmentation and $+3.9$ mAP for vectorized detection on nuScenes validation (Xiong et al., 2023).

Two other systems address related persistence and fusion problems. GNMap is an end-to-end generative neural network that fuses multiple vectorized tiles produced across several tours into a single globalized HD map tile under world coordinates. It uses a multi-layer and attention-based autoencoder with a self-supervised pretraining stage for completeness and a supervised finetuning stage for semantic correctness, and reports an F1 score of 74.0, more than 5% above the best single-tour baselines in its evaluation (Fan et al., 2024). NeMO, by contrast, maintains a persistent scene-level BEV “big map”

$$\mathcal{F}^{map} \subseteq \mathbb{R}_{O_g}^{H_{map} \times W_{map} \times K},$$

initialized to zeros and updated by shared-weight fusion at each grid cell as the vehicle moves (Zhu et al., 2023). The paper reports consistent gains over overwrite and max-pooling baselines in scene-level BEV map generation on both nuScenes and BDD-Map (Zhu et al., 2023).

The localization problem can also be cast in a “NeoMap”-like way. “MapLocNet” performs coarse-to-fine neural feature registration between rasterized navigation maps and visual BEV features, estimating a 3-DoF pose correction on top of noisy GPS (Wu et al., 2024). It reports nearly 10% and 20% localization-accuracy improvements over OrienterNet on nuScenes and Argoverse, together with 30 and 16 FPS improvement on single-view and surround-view settings, respectively (Wu et al., 2024). A plausible implication is that low-cost navigation maps and neural registration can serve as a practical complement to HD-map-centric localization pipelines.

3. Graph correspondence, scale normalization, and landmark embedding

A second major theme is map alignment by graph correspondence rather than by dense geometry. “UM3: Unsupervised Map to Map Matching” models each map as an undirected graph $G = (\mathcal{V}, \mathcal{E})$, where nodes are geographic locations with latitude–longitude coordinates and edges are road segments (Ying et al., 23 Aug 2025). Its correspondence matrix $\mathbf{S} \in [0,1]^{n_s \times n_t}$ is learned without any matched supervision. The framework introduces pseudo coordinates by min–max normalization within each map: $$\begin{aligned} \bar{x}^{(1)}(v) &= \frac{\mathrm{lat}(v) - \mathrm{lat}_{\min}}{\mathrm{lat}_{\max} - \mathrm{lat}_{\min}},\ \bar{x}^{(2)}(v) &= \frac{\mathrm{lon}(v) - \mathrm{lon}_{\min}}{\mathrm{lon}_{\max} - \mathrm{lon}_{\min}}. \end{aligned}$$ These pseudo coordinates capture relative spatial layout and support scale-invariant learning (Ying et al., 23 Aug 2025). Node embeddings are produced by a shared 3-layer GCN; feature similarity is fused with a learned geometric kernel and normalized by Sinkhorn; and the unsupervised objective combines distance consistency and local structural consistency: $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$0 On Boston, Ichikawa, Shanghai, and Bremen, the method surpasses ICP, Sinkhorn distance, HMM-based baselines, and supervised NGM, with particularly large gains in high-noise and large-scale settings (Ying et al., 23 Aug 2025).

The landmark-embedding line represented by NeuMaps addresses a related but not identical problem: extracting a geometry-aware embedding from a selected subset of vertices. Given a landmark set $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$1 inside a full data graph $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$2, NeuMaps constructs a Neumann Laplacian

$$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$3

on the landmark subgraph and solves a generalized eigenproblem

$$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$4

to obtain the embedding coordinates (Sule et al., 10 Feb 2025). The resulting map is an embedding of the reflecting random walk on the landmark subgraph, and the discrete Neumann boundary condition yields a built-in Nyström extension for non-landmark points (Sule et al., 10 Feb 2025). On UCI digits, NeuMaps improves over Roseland from NMI $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$5, ACC $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$6 to NMI $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$7, ACC $$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$8; in molecular dynamics of butane, it recovers the dihedral angle more accurately and remains stable when significant points are removed (Sule et al., 10 Feb 2025).

Taken together, these works show two complementary strategies for “NeoMap”-style correspondence. UM3 treats maps as explicit graphs requiring node-to-node matching under noise and partial heterogeneity (Ying et al., 23 Aug 2025). NeuMaps treats distinguished samples as a subgraph whose embedding should preserve a diffusion geometry under Neumann boundary conditions (Sule et al., 10 Feb 2025). This suggests that map learning in graph domains often alternates between two tasks: estimating correspondences and defining an intrinsic coordinate system.

4. Surface maps and mathematical map presentations

Outside geographic mapping, “NeoMap” also names or motivates systems for maps between manifolds and branched coverings. “Neural Surface Maps” defines a neural surface map as a neural network

$$\begin{aligned} z_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_z)) \ r_t &= \sigma(\mathrm{Conv2D}([p_{t-1}, o], w_r)) \ \tilde{p}_t &= \tanh(\mathrm{Conv2D}([r_t \odot p_{t-1}, o], w_h)) \ p_t &= (1 - z_t) \odot p_{t-1} + z_t \odot \tilde{p}_t . \end{aligned}$$9

with $+4.32$0 a canonical planar domain and $+4.32$1 interpreted as a differentiable surface map whose image is a 2-manifold in $+4.32$2 (Morreale et al., 2021). Surfaces are represented by overfitting a neural chart to a UV parameterization; reparameterizations and surface-to-surface maps are represented by additional networks $+4.32$3; and distortion objectives are written directly in terms of Jacobians and metric tensors. The paper uses symmetric Dirichlet and conformal distortion energies and optimizes them with automatic differentiation, without combinatorial bookkeeping over meshes (Morreale et al., 2021). The method demonstrates parameterization optimization, pairwise surface correspondences, and cycle-consistent mapping across collections, all through compositions of neural maps.

A more classical mathematical use appears in “Presentations of NET maps”. There, “NET maps” are nearly Euclidean Thurston maps: branched coverings $+4.32$4 whose critical points are simple and whose postcritical set has exactly four points (Floyd et al., 2017). The paper proves that every NET map is Thurston equivalent to one given by a presentation diagram determined by a $+4.32$5 integer matrix $+4.32$6 with $+4.32$7, a translation term $+4.32$8, and a set of green geodesic arcs encoding a push map (Floyd et al., 2017). This produces a normal form from simple affine data, supports explicit modular-group actions by matrix twisting and translations, and underlies the NETmap computational system and census described by the authors (Floyd et al., 2017).

These two strands differ profoundly in domain and method. Neural Surface Maps replace discrete surface correspondences by composable neural functions over a canonical domain (Morreale et al., 2021). NET-map presentations reduce a class of topological branched coverings to affine normal forms and push maps (Floyd et al., 2017). Their coexistence under the broader “NeoMap” framing underscores a general property of the term in research usage: “map” is interpreted not only as a geographic artifact but also as a morphism, parameterization, or equivalence class.

5. Time-aware synoptic maps and encyclopedic visualization

“Project G.N.O.S.I.S.: Geographical Network Of Synoptic Information System” supplies a non-neural but conceptually important antecedent for “NeoMap” as an encyclopedic map interface (Oliva, 2012). Its core idea is to add the missing dimension of time to a global map platform such as Google Earth, allowing users to “browse the map of the Earth in time” and to “dress up” the globe with “skin-maps” representing geology, genetics, agriculture, ethnology, linguistics, musicology, metallurgy, epidemiology, religion, dialects, and other histories (Oliva, 2012). The proposal explicitly states that the area for each event is function of time and eventually vanishes when the event has ceased its effect, and that active links over each area should redirect the viewer to the corresponding source, such as a Wikipedia page (Oliva, 2012).

The project anticipates several design motifs that later reappear in neural mapping systems. First, it treats the globe as a rendering substrate over which multiple time-indexed semantic layers are overlaid. Second, it assumes progressive disclosure: by zooming the area also the level of definition of the information will increase (Oliva, 2012). Third, it proposes collaborative curation through a Wikipedia-like model and version control through software revision systems such as Apache Subversion (Oliva, 2012).

No algorithmic framework is provided, and the paper is explicitly high-level. Nevertheless, its emphasis on spatiotemporal layers, interactive retrieval, and encyclopedic breadth suggests a precursor to later “NeoMap”-style systems that store and update maps as information-bearing layers rather than as static cartographic products.

6. NeoMap as training-free novel-view synthesis

The most literal current use of the name is “NeoMap: Training-free Novel-View Synthesis from Single Images and Videos” (Li et al., 2 Jul 2026). Here NeoMap is not a map-construction system in the cartographic sense but a training-free framework for locating view-consistent solutions inside the natural video manifold of a pre-trained video model. The input is either a single RGB image or a monocular video plus a desired camera trajectory; the output is a novel-view video that should be photorealistic, temporally coherent, and geometrically consistent with the source view and target cameras (Li et al., 2 Jul 2026).

The method begins by estimating depth and camera pose, then warping source content into the target views to form a warped prior video $+4.32$9 with visibility masks $+5.02$0. This prior is encoded by a VAE to a latent $+5.02$1, then noised to a near-noise state $+5.02$2. NeoMap’s key mechanism is convergent manifold alternating projection iterations between two sets: the model’s natural video manifold $+5.02$3 and the set $+5.02$4 of latents whose decoded visible pixels match the warped prior (Li et al., 2 Jul 2026). The method alternates an Anchored Manifold Projection, which uses the flow vector field to move toward $+5.02$5 while feature-wise anchoring reliable latent regions to the warped prior, with a Pixel-Constrained Projection, which enforces the visible pixels exactly: $+5.02$6 The alternating scheme seeks a point in $+5.02$7, after which standard unguided sampling from the optimized near-noise state produces the final video (Li et al., 2 Jul 2026).

Across Tanks-and-Temples, LLFF, and DAVIS, NeoMap is reported to outperform camera-conditioned, warping-and-inpainting, and noise-manipulation baselines in visual fidelity and view consistency (Li et al., 2 Jul 2026). On Tanks-and-Temples, it achieves PSNR $+5.02$8, LPIPS $+5.02$9, FID-192 $+5.53$0, and CLIP-S $+5.53$1; on DAVIS, it achieves FID-192 $+5.53$2, FVD $+5.53$3, ATE $+5.53$4, RPE-T $+5.53$5, and RPE-R $+5.53$6 (Li et al., 2 Jul 2026). This usage broadens the semantic range of “NeoMap”: the “map” is a manifold-constrained mapping from optimized noise to a view-consistent video rather than a spatial database.

7. Conceptual synthesis and recurring design patterns

Across these papers, several recurrent design patterns define the broader “NeoMap” idea. One is the replacement of explicit discrete map entities by latent or neural state. This is explicit in global neural map priors $+5.53$7 for HD maps, scene-level BEV big maps $+5.53$8, surface charts $+5.53$9, and video-manifold latents $+3.9$0 (Xiong et al., 2023, Zhu et al., 2023, Morreale et al., 2021, Li et al., 2 Jul 2026). Another is the use of structured priors to fuse incomplete evidence: graph topology in MfM and UM3, local connectivity in NeuMaps, world-coordinate tile fusion in GNMap, current-to-prior attention in NMP, and pixel-constrained warped priors in NeoMap NVS (Toso et al., 2024, Ying et al., 23 Aug 2025, Sule et al., 10 Feb 2025, Fan et al., 2024, Xiong et al., 2023, Li et al., 2 Jul 2026).

A further commonality is that many systems are designed for regimes where direct geometric optimization is brittle, incomplete, or expensive. MfM is intended for sparse uncalibrated street imagery where feature matching is unreliable (Toso et al., 2024). NMP, GNMap, and NeMO target persistent map construction from partial traversals and sensor-limited perception (Xiong et al., 2023, Fan et al., 2024, Zhu et al., 2023). UM3 avoids labeled correspondences in large-scale map alignment (Ying et al., 23 Aug 2025). NeoMap NVS avoids task-specific finetuning by searching for a valid initial noise on a pre-trained video manifold (Li et al., 2 Jul 2026).

At the same time, the term remains heterogeneous. In one setting it denotes a concrete 2026 method for training-free novel-view synthesis (Li et al., 2 Jul 2026). In others it acts as a design label for automatic map construction, neural priors, map localization, manifold parameterization, or synoptic globe interfaces (Toso et al., 2024, Xiong et al., 2023, Wu et al., 2024, Morreale et al., 2021, Oliva, 2012). The literature therefore does not support a single canonical definition. A more accurate characterization is that “NeoMap” marks a research tendency: maps are treated as learned, compositional, and updateable intermediates that mediate between observation, structure, and inference.