A-TDOM: Adaptive Token Access and Digital Mapping

• A-TDOM is a multifaceted term that denotes both adaptive token-domain multiple access in semantic communications and active true digital orthophoto map generation in remote sensing.
• In semantic communications, A-TDOM employs a transformer-based scheduler that jointly manages token selection, semantic interference, and channel-aware power allocation.
• In geospatial mapping, A-TDOM utilizes on-the-fly structure-from-motion and 3D Gaussian splatting to generate near real-time, distortion-free orthophotos with enhanced rendering quality.

Searching arXiv for the cited A-TDOM-related papers and adjacent work to ground the article. “A-TDOM” is an overloaded designation used in several technically distinct research contexts. In recent literature, it most prominently denotes Adaptive Token-Domain Multiple Access, the token-domain scheduling component within the ATS-ToDMA framework for cross-modal semantic communications (Kadam et al., 3 Jul 2026). In geospatial mapping, the same string denotes Active True Digital Orthophoto Map, a near real-time TDOM generation method based on On-the-Fly 3DGS optimization (Xu et al., 16 Sep 2025). Related usages include an informal variant of ATDOC, the Auxiliary Target Domain-Oriented Classifier for domain adaptation (Liang et al., 2020), and, in multi-agent reinforcement learning, an informal shorthand for the actor-critic instantiation of a Time Dynamical Opponent Model, named TDOM-AC in the original paper (Tian et al., 2022). Because these meanings are unrelated at the methodological level, precise interpretation depends on disciplinary context.

1. Semantic communications usage: Adaptive Token-Domain Multiple Access

In semantic communications, A-TDOM refers to Adaptive Token-Domain Multiple Access, the token-domain multiple-access scheduling mechanism embedded in ATS-ToDMA, a cross-modal semantic communication framework spanning text, image, and speech modalities (Kadam et al., 3 Jul 2026). ATS-ToDMA performs three functions jointly: adaptive token selection (ATS), token scheduling via ToDMA while mitigating intra-modal and cross-modal semantic interference, and semantic-aware transmit-power allocation. A-TDOM is specifically the ToDMA scheduling component.

The framework treats semantic tokens as schedulable transmission units. Each user $u$ and modality $m \in \{\text{text}, \text{image}, \text{speech}\}$ produces tokens through a modality-specific encoder such as an LLM, ViT, or speech encoder, with token embeddings $e_i \in \mathbb{R}^d$ normalized so that $\|e_i\|=1$ (Kadam et al., 3 Jul 2026). ATS reduces the token set by thresholding token-importance scores $s_i=q(t_i)$, retaining a token $t_i$ when $s_i > \tau_{\text{ATS}}$. The remaining tokens are assigned to token-domain transmission slots, where multiple tokens may be transmitted simultaneously.

A-TDOM is motivated by the observation that simultaneous transmission of semantically similar tokens induces semantic interference. The scheduler is therefore designed to account jointly for token importance, semantic similarity, channel conditions, and feasibility constraints. This differs from signal-centric orthogonal and non-orthogonal access schemes by elevating semantic-token interactions to first-class scheduling variables (Kadam et al., 3 Jul 2026).

2. System model, SSINR, and interference-aware scheduling

The semantic-communication formulation defines token similarity by cosine similarity,

$$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$

and declares a pair semantically similar when $\xi_{ij} > \overline{\gamma}$, with indicator $\mathbbm{1}_{ij}$ (Kadam et al., 3 Jul 2026). Semantic interference is modality dependent: intra-modal pairs induce stronger effective distortion than cross-modal pairs, with $m \in \{\text{text}, \text{image}, \text{speech}\}$0.

The central design metric is the Semantic Signal-to-Interference-plus-Noise Ratio (SSINR),

$m \in \{\text{text}, \text{image}, \text{speech}\}$1

where $m \in \{\text{text}, \text{image}, \text{speech}\}$2 is the token transmit power, $m \in \{\text{text}, \text{image}, \text{speech}\}$3 is thermal noise power, $m \in \{\text{text}, \text{image}, \text{speech}\}$4 is a product-power term scaled by $m \in \{\text{text}, \text{image}, \text{speech}\}$5, and $m \in \{\text{text}, \text{image}, \text{speech}\}$6 is a channel-aware gating vector derived from CSI and SNR through

$m \in \{\text{text}, \text{image}, \text{speech}\}$7

The SSINR combines channel impairments and semantic interference in a single reliability metric (Kadam et al., 3 Jul 2026).

The corresponding joint optimization problem maximizes semantic throughput subject to token-selection consistency, per-slot occupancy, expected and instantaneous semantic-interference constraints, and per-token SSINR reliability constraints. The scheduling variable $m \in \{\text{text}, \text{image}, \text{speech}\}$8 indicates token-to-slot assignment, with $m \in \{\text{text}, \text{image}, \text{speech}\}$9 ensuring that a token is assigned only when selected, and $e_i \in \mathbb{R}^d$0 enforcing feasible occupancy per slot (Kadam et al., 3 Jul 2026).

A-TDOM is implemented as a transformer-based scheduler. Tokens are represented as $e_i \in \mathbb{R}^d$1, contextualized by a transformer encoder,

$e_i \in \mathbb{R}^d$2

with standard self-attention

$e_i \in \mathbb{R}^d$3

The output is a slot-assignment distribution $e_i \in \mathbb{R}^d$4, and inference uses

$e_i \in \mathbb{R}^d$5

Training enforces soft occupancy and interference constraints through a loss containing penalties on expected slot occupancy $e_i \in \mathbb{R}^d$6, expected per-slot interference

$e_i \in \mathbb{R}^d$7

and expected semantic interference $e_i \in \mathbb{R}^d$8 (Kadam et al., 3 Jul 2026). During inference, hard constraints are enforced by pruning low-$e_i \in \mathbb{R}^d$9 tokens when $\|e_i\|=1$0 exceeds $\|e_i\|=1$1, and by iteratively removing tokens with the highest interference contribution when $\|e_i\|=1$2.

3. Analytical bounds, power allocation, and reported performance

The ATS-ToDMA paper derives three analytical results directly relevant to A-TDOM (Kadam et al., 3 Jul 2026). First, under bounded similarity and distortion coefficients, the total interference for $\|e_i\|=1$3 equal-power tokens in a slot satisfies

$\|e_i\|=1$4

Second, under the same assumptions and the reliability requirement $\|e_i\|=1$5, the maximum feasible slot occupancy is

$\|e_i\|=1$6

Third, the semantic-aware power-allocation subproblem admits a closed-form approximation. With coupling weights

$\|e_i\|=1$7

the approximate power allocation is

$\|e_i\|=1$8

obtained via a first-order Neumann approximation when the coupling matrix has spectral radius $\|e_i\|=1$9.

The reported simulation setup uses $s_i=q(t_i)$0 users, modalities text/image/speech, embedding dimension $s_i=q(t_i)$1, Rayleigh fading with AWGN, $s_i=q(t_i)$2, calibrated coefficients $s_i=q(t_i)$3 and $s_i=q(t_i)$4, ATS threshold $s_i=q(t_i)$5, similarity cap $s_i=q(t_i)$6, target SSINR $s_i=q(t_i)$7, and 1000 Monte Carlo trials (Kadam et al., 3 Jul 2026). Benchmarks include OMA, Semantic NOMA, Random-TS, Greedy ATS, and Equal-Power.

Semantic throughput is evaluated by

$s_i=q(t_i)$8

ATS-ToDMA is reported to achieve the highest semantic throughput across user loads relative to OMA, Semantic NOMA, and Random-TS. Against Greedy ATS, the paper reports throughput $s_i=q(t_i)$9 versus $t_i$0 bits/s/Hz, semantic accuracy $t_i$1 versus $t_i$2, aggregate semantic interference $t_i$3 versus $t_i$4, average SSINR $t_i$5 versus $t_i$6, and average transmit power $t_i$7 versus $t_i$8 W (Kadam et al., 3 Jul 2026). The paper also reports that the analytical interference bound, feasible occupancy bound, and closed-form power allocation track simulation and SCA-based optimization under the tested regimes.

This suggests that, within semantic communications, A-TDOM is not merely a slot allocator but a constrained resource-management layer coupling semantic similarity, modality interactions, and channel-aware reliability.

4. Geospatial mapping usage: Active True Digital Orthophoto Map

In remote sensing and photogrammetry, A-TDOM denotes Active True Digital Orthophoto Map, a near real-time TDOM generation method based on On-the-Fly SfM and streaming 3D Gaussian Splatting (Xu et al., 16 Sep 2025). Here the term is unrelated to semantic communications.

A TDOM differs from a conventional DOM in that it removes relief displacement and building lean, preserves true vertical geometry, handles occlusions, and represents only the topmost visible surfaces such as roofs and ground (Xu et al., 16 Sep 2025, Wang et al., 2024). Traditional TDOM pipelines depend on accurate camera poses and DSMs, use Z-buffering or ray-casting for occlusion detection, and typically perform global SfM, dense reconstruction, and orthorectification only after all images have been collected, introducing latency (Xu et al., 16 Sep 2025).

A-TDOM addresses these constraints by maintaining an online 3DGS scene that is updated as images stream in. Each new image triggers five stages: On-the-Fly SfM for pose estimation and sparse-point updates; key-region masking via reprojected sparse points and Delaunay triangulation; Gaussian sampling and integration in previously unseen or coarsely reconstructed regions based on a gradient discrepancy map; adaptive 3DGS optimization; and orthogonal splatting to render an updated TDOM immediately (Xu et al., 16 Sep 2025).

The method initializes 3DGS with 2000 iterations on a seed image set. For each incoming image $t_i$9, it estimates pose $s_i > \tau_{\text{ATS}}$0 via 2D–3D matches and PnP, updates a local bundle adjustment window, renders the current 3DGS in the masked region $s_i > \tau_{\text{ATS}}$1, computes LoG gradients on the image and rendering, and defines a discrepancy region

$s_i > \tau_{\text{ATS}}$2

with $s_i > \tau_{\text{ATS}}$3 (Xu et al., 16 Sep 2025). New Gaussians are sampled within masked triangles and initialized through barycentric interpolation of 3D positions and colors.

The online Gaussian field is rendered by orthographic splatting rather than perspective rendering. For projected screen point $s_i > \tau_{\text{ATS}}$4, the splat weight is

$s_i > \tau_{\text{ATS}}$5

and front-to-back alpha compositing along an orthographic ray uses

$s_i > \tau_{\text{ATS}}$6

This handles visibility without an explicit DSM, unlike classical TDOM workflows (Xu et al., 16 Sep 2025). The orthographic projection matrix $s_i > \tau_{\text{ATS}}$7 and the covariance transform $s_i > \tau_{\text{ATS}}$8 implement true ortho rendering.

The reported performance is explicitly near real-time rather than fully real-time. Across the phantom3-npu, phantom3-ieu, WHU-S, and WHU-Q1 datasets, measured throughput is $s_i > \tau_{\text{ATS}}$9, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$0, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$1, and $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$2 FPS, respectively (Xu et al., 16 Sep 2025). Optimization time from the end of pose estimation to completed TDOM is reported as $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$3, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$4, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$5, and $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$6 seconds for A-TDOM, compared with $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$7, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$8, $$\xi_{ij} = \frac{\langle t_i, t_j \rangle}{\|t_i\|\|t_j\|},$$9, and $\xi_{ij} > \overline{\gamma}$0 seconds for Tortho-Gaussian, and $\xi_{ij} > \overline{\gamma}$1, $\xi_{ij} > \overline{\gamma}$2, $\xi_{ij} > \overline{\gamma}$3, and $\xi_{ij} > \overline{\gamma}$4 seconds for AdR-Gaussian (Xu et al., 16 Sep 2025). The paper reports sharper building edges, removal of building facades, and continuity of slender structures in qualitative comparisons.

This geospatial usage should be read alongside related TDOM-generation work such as “Tortho-Gaussian” (Wang et al., 2024) and 2DGS-based orthoimage generation (Wang et al., 25 Mar 2025). Those papers do not use the term A-TDOM in the same formal sense, but they establish the surrounding methodological landscape: orthographic Gaussian splatting, elimination of DSM dependence, divide-and-conquer scalability, and improved rendering of weak-texture and slender structures.

5. Other uses of the string: ATDOC and TDOM-AC

The string “A-TDOM” also appears informally in other literatures, but these are naming variants rather than canonical method names.

In domain adaptation, the official method name is ATDOC, short for Auxiliary Target Domain-Oriented Classifier (Liang et al., 2020). The paper notes that “A-TDOM” may appear informally as a naming variant meaning “Auxiliary Target Domain-Oriented [Model/Memory/Classifier],” but the official terminology is ATDOC. The method introduces a non-parametric auxiliary classifier $\xi_{ij} > \overline{\gamma}$5 operating only on target data to improve pseudo labels and reduce source bias. Two variants are defined: ATDOC-NC, based on target class centroids updated via an EMA memory, and ATDOC-NA, based on neighborhood aggregation in a target-feature memory bank (Liang et al., 2020). Because the source paper does not use A-TDOM as the formal method name, equating A-TDOM with ATDOC should be understood as informal usage.

In multi-agent reinforcement learning, the official term is TDOM-AC, short for Multi-Agent Actor-Critic with Time Dynamical Opponent Model (Tian et al., 2022). Here “A-TDOM” can be used informally to refer to the actor-critic instantiation, but the paper itself does not name the method A-TDOM. TDOM models opponent policies as time-evolving through a learning/improvement dynamic and yields a lower bound involving a KL term between the learned opponent model and the opponents’ marginal policy (Tian et al., 2022). The actor conditions on predicted opponent actions, the critic is centralized, and the opponent model is trained with an entropy-regularized objective. Reported average per-update running times are $\xi_{ij} > \overline{\gamma}$6 s for TDOM-AC, $\xi_{ij} > \overline{\gamma}$7 s for ROMMEO, and $\xi_{ij} > \overline{\gamma}$8 s for PR2, with favorable stability and convergence in cooperative and mixed cooperative-competitive settings (Tian et al., 2022).

These two cases illustrate that the string “A-TDOM” is best treated as context dependent. In formal citation practice, ATDOC and TDOM-AC should retain their original names.

6. Disambiguation, misconceptions, and research significance

A common misconception is that A-TDOM refers to a single method or research line. The literature instead supports at least four distinct interpretations:

Usage context	Expansion	Formal name in source
Semantic communications	Adaptive Token-Domain Multiple Access	A-TDOM within ATS-ToDMA
Geospatial mapping	Active True Digital Orthophoto Map	A-TDOM
Domain adaptation	Auxiliary Target Domain-Oriented Classifier	ATDOC
Multi-agent RL	Actor-Critic with Time Dynamical Opponent Model	TDOM-AC

The semantic-communications and geospatial usages are both formal and current, but they are methodologically unrelated. The former concerns cross-modal semantic-token scheduling under semantic interference and channel constraints (Kadam et al., 3 Jul 2026); the latter concerns streaming UAV photogrammetry and orthographic rendering with online 3DGS updates (Xu et al., 16 Sep 2025). The ATDOC and TDOM-AC cases are better understood as adjacent naming ambiguities rather than direct synonyms.

Another misconception is that all “A-TDOM” usages involve orthophoto mapping because TDOM is a long-established abbreviation for True Digital Orthophoto Map. The recent ATS-ToDMA work demonstrates that, in communications, the same letter sequence has been reassigned to token-domain multiple access (Kadam et al., 3 Jul 2026). Conversely, in the geospatial literature, A-TDOM inherits the orthophoto meaning of TDOM and emphasizes “active” or online map updating (Xu et al., 16 Sep 2025).

From a broader research perspective, the coexistence of these meanings reflects a wider trend toward task-structured resource allocation under streaming constraints. In semantic communications, tokens, interference, and power are optimized as semantically meaningful units (Kadam et al., 3 Jul 2026). In active TDOM generation, Gaussians, overlap masks, and orthographic rendering are updated as new sensory evidence arrives (Xu et al., 16 Sep 2025). This suggests a shared systems-level pattern—online selection, structured scheduling, and constraint-aware inference—even though the mathematical objects and application domains differ.

7. Outlook

Within semantic communications, A-TDOM points toward denser multi-user, cross-modal settings in which semantic similarity becomes a resource-allocation variable rather than a nuisance parameter. The ATS-ToDMA paper identifies future integration with foundation-model-based encoders, edge intelligence, distributed scheduling, dynamic network conditions, and 6G SemCom standardization as natural directions (Kadam et al., 3 Jul 2026).

Within photogrammetry and mapping, A-TDOM is part of a shift from offline DSM-dependent orthophoto production toward online Gaussian-splatting-based scene modeling and rendering (Xu et al., 16 Sep 2025). Related work on orthographic 3DGS and 2DGS indicates continuing interest in eliminating explicit DSM construction, improving weak-texture and slender-structure rendering, and scaling to large urban scenes via divide-and-conquer training and tiled orthographic rasterization (Wang et al., 2024, Wang et al., 25 Mar 2025). A plausible implication is that active TDOM generation will increasingly be coupled to real-time downstream tasks such as geo-localization, where TDOM and DSM layers can serve as lightweight orthographic reference substrates, as in PiLoT v2 (Liu et al., 30 Jun 2026).

Accordingly, the most precise encyclopedia treatment of “A-TDOM” is not a single-method entry but a disambiguated technical term. In current arXiv usage, its primary formal meanings are Adaptive Token-Domain Multiple Access in semantic communications (Kadam et al., 3 Jul 2026) and Active True Digital Orthophoto Map in near real-time geospatial mapping (Xu et al., 16 Sep 2025), with additional informal overlap involving ATDOC (Liang et al., 2020) and TDOM-AC (Tian et al., 2022).