Domain-MoT: Cross-Domain Tracking

• Domain-MoT is a multifaceted concept defining methods for adapting multi-object tracking across diverse domains in computer vision, cold-atom physics, and electromagnetics.
• It details adaptive strategies such as self-training, test-time adaptation, and language-based supervision to mitigate domain shifts and improve metrics like HOTA and MOTA.
• The framework also addresses sensor-specific and platform-specific challenges, offering solutions that enhance numerical stability and robust performance across various applications.

In current research usage, “Domain-MoT” spans several technically distinct literatures. A major computer-vision sense concerns multi-object tracking (MOT) under domain shift, where models trained in one visual domain must generalize or adapt to another; related work extends this setting to thermal, drone, autonomous-driving, and robotic platforms. The same lexical form also appears in cold-atom physics, where a membrane-defined magneto-optical trap localizes atoms in a sub-millimeter hole, and in computational electromagnetics, where MOT denotes a marching-on-in-time solution of the time-domain PMCHWT equation (Wang et al., 2022, Yu et al., 2022, Lee et al., 2020, Le et al., 18 Sep 2025).

1. Cross-domain MOT as the central vision interpretation

Within computer vision, Domain-MoT is most naturally associated with the observation that most MOT systems are trained and validated in-domain, even though deployment is inherently cross-domain. The domain shift can involve crowd density, illumination, viewpoint, background, resolution, indoor/outdoor setting, motion pattern, sensor modality, or the synthetic-to-real gap. In query-based trackers and tracking-by-detection pipelines alike, this shift affects both detection quality and identity association, so performance degradation is typically visible in HOTA, AssA, DetA, MOTA, IDF1, and IDS (Yu et al., 2022, Segu et al., 2023).

This framing is explicit in work that treats MOT17 and MOT20 as distinct domains rather than merely different splits. “Generalizing Multiple Object Tracking to Unseen Domains by Introducing Natural Language Representation” defines a cross-domain benchmark by training on MOT17 and validating or testing on MOT20, and reports that existing trackers such as TransTrack, TraDeS, CenterTrack, and MOTR degrade sharply under transfer (Yu et al., 2022). “DARTH: Holistic Test-time Adaptation for Multiple Object Tracking” broadens the notion of domain shift to include sim-to-real, outdoor-to-indoor, and indoor-to-outdoor transfer, with scenarios such as SHIFT $\rightarrow$ BDD100K and MOT17 $\rightarrow$ DanceTrack (Segu et al., 2023).

A recurrent technical point is that MOT is not reducible to detection alone. In appearance-based tracking, domain shift damages the detector, the appearance embedding space, and the temporal consistency required for association. This is why methods in this area either adapt several subsystems jointly or inject representations that are claimed to be more domain invariant than raw visual features (Segu et al., 2023, Yu et al., 2022).

2. Adaptation strategies for cross-domain tracking

One adaptation line is self-training on unlabeled target data. “PieTrack: An MOT solution based on synthetic data training and self-supervised domain adaptation” trains YOLOX-X from scratch on MOTSynth, then performs iterative pseudo-label refinement on the unlabeled MOT17 training split. The paper specifies an initial pseudo-label confidence threshold $T = 0.5$, subsequent threshold $T = 0.1$, initial fine-tuning length $N = 40$, and later fine-tuning length $N = 20$. On the MOT17 training set, the method improves from a baseline HOTA of 52.87 to 55.66 at Iter.1 and 56.82 at Iter.2, while FN drops from 33258 to 28643 and 28569; with multi-scale ensemble inference, HOTA rises to 57.68, and the final test result is HOTA 58.7 (Wang et al., 2022). The same paper is explicit that the method does not use an adversarial alignment loss, a teacher-student consistency loss, or pre-trained weights; adaptation is realized through iterative pseudo-label generation and fine-tuning (Wang et al., 2022).

A second line is holistic test-time adaptation. DARTH uses a teacher-student framework with EMA update $\xi \leftarrow \tau \xi + (1-\tau)\theta$ and $\tau = 0.998$, combining detection consistency at the RPN and RoI stages with a patch contrastive loss for appearance adaptation (Segu et al., 2023). The paper argues that adapting only the detector is insufficient for MOT. Quantitatively, in SHIFT $\rightarrow$ BDD100K, No Adaptation gives DetA 12.0, MOTA -66.4, HOTA 17.3, IDF1 18.5, and AssA 28.9, whereas DARTH reports DetA 15.2, MOTA 8.3, HOTA 20.6, IDF1 23.7, and AssA 33.1; in MOT17 $\rightarrow$ DanceTrack, the method improves HOTA from 21.5 to 31.6 and AssA from 9.0 to 17.7 (Segu et al., 2023).

A third line introduces language-like supervision. LTrack augments MOTR with Visual Context Prompting (VCP) and Visual-Language Mixing (VLM), using a Trackbook of 56 description phrases and a frozen CLIP text encoder to generate pseudo textual descriptions that are treated as domain invariant high-level summaries (Yu et al., 2022). In the MOT17 $\rightarrow$0 MOT20 setting, LTrack reports HOTA 46.8, AssA 45.4, DetA 48.4, MOTA 57.8, IDF1 61.1, and IDS 1841, compared with MOTR at HOTA 43.7 and IDF1 56.8 (Yu et al., 2022). IP-MOT generalizes this line by replacing coarse handcrafted textual descriptions with instance-level pseudo textual descriptions obtained through CLIP-based prompt tuning, and by introducing a query-balanced strategy plus a deduplication module (Luo et al., 2024). In the same MOT17 $\rightarrow$1 MOT20 protocol, IP-MOT reports MOTA 68.3, IDF1 62.5, HOTA 49.2, AssA 44.6, and DetA 55.3, exceeding the previous cross-domain language-based tracker LTrack at HOTA 46.2 and IDF1 60.4 (Luo et al., 2024).

These results support two recurring corrections to common assumptions. First, cross-domain MOT need not rely on explicit adversarial domain alignment; self-training and test-time adaptation can be sufficient in practice (Wang et al., 2022, Segu et al., 2023). Second, natural-language priors are not treated as literal human captions of every tracked instance; both LTrack and IP-MOT construct pseudo textual descriptions, but IP-MOT argues that coarse class-level prompts are not sufficiently discriminative for instance association (Yu et al., 2022, Luo et al., 2024).

3. Sensor-specific and platform-specific domains

A distinct but related use of Domain-MoT concerns deployment domains defined by sensing modality or platform. In thermal imaging, “Enhancing Thermal MOT” modifies the association stage of ByteTrack and OCSORT by combining motion similarity with thermal identity similarity computed from normalized histogram comparisons inside thermal ROIs. The fused score is

$\rightarrow$2

The paper reports the best trade-off at $\rightarrow$3 for ByteTrack and $\rightarrow$4 for OCSORT, with ByteTrack on thermal sequences improving from IDF1 62.6% and MOTA 65.5% to IDF1 63.8% and MOTA 66.4%, and OCSORT improving from IDF1 57.8% and MOTA 54.4% to IDF1 58.6% and MOTA 56.4% (Ahmar et al., 2024). The same work introduces an RGB-Thermal MOT dataset with 30 sequences and 9000 frames per modality total, collected at 5 urban intersections (Ahmar et al., 2024).

In autonomous driving, MotionTrack establishes an end-to-end transformer baseline for multi-class MOT with LiDAR-only and LiDAR-camera fusion inputs. Its data-association module uses transformer-based feature refinement but performs matching through a raw dot-product cost matrix rather than softmax attention, because the paper reports gradient-vanishing issues when softmax is used directly for association (Zhang et al., 2023). On nuScenes, MotionTrack-Voxel-LC reaches AMOTA 0.55; the ablation also attributes a large fraction of the gain to the transformer-based DA module, with MotionTrack-Voxel-L improving from AMOTA 0.22 without the transformer to 0.62 with it, and MotionTrack-Voxel-LC improving from 0.23 to 0.69 in the corresponding ablation (Zhang et al., 2023).

In drone tracking, DroneMOT argues that classical static-camera assumptions break down because both the camera platform and the objects move. It therefore combines Dual-domain Integrated Attention (DIA) on the detection side with Motion-Driven Association (MDA), whose two core components are Adaptive Feature Synchronization (AFS) and Dual Motion-based Prediction (DMP) (Wang et al., 2024). On VisDrone2019-MOT, DroneMOT reports IDF1 58.6, MOTA 43.7, and IDs 1112; on UAVDT, it reports IDF1 69.6, MOTA 50.1, and IDs 129 (Wang et al., 2024).

In robotics, DepTR-MOT adds instance-level depth prediction to a DETR-based detector and uses depth-informed trajectory refinement to improve association under occlusion and close-proximity interactions. The method uses foundation model-based instance-level soft depth label supervision, dense depth map distillation for global depth consistency, and a depth-aware association cost $\rightarrow$5 (Deng et al., 22 Sep 2025). It reports HOTA 27.59 on QuadTrack and 44.47 on DanceTrack, and emphasizes stronger gains on the robotic QuadTrack setting, where IDF1 improves by +4.96, MOTA by +5.72, and AssA by +6.35 (Deng et al., 22 Sep 2025).

4. Representation and association as domain-robust design variables

Some work addresses Domain-MoT not by explicit adaptation, but by strengthening the internal representation used for long-horizon association. “MOT FCG++: Enhanced Representation of Spatio-temporal Motion and Appearance Features” preserves the hierarchical clustering framework of MOT_FCG while replacing its weaker trajectory summaries with Diagonal Modulated GIoU (DGIoU), Average Constant Velocity Modeling (ACV), and Dynamic Appearance (DA) (Fang, 2024).

The motion-side change is motivated by the claim that IoU alone does not capture relative geometry adequately for clustering association. In the reported ablation on MOT17-train, IoU yields HOTA 77.06, MOTA 87.19, IDF1 86.09, AssA 76.76, and IDSW 2262, whereas DGIoU yields HOTA 77.70, MOTA 87.30, IDF1 87.06, AssA 77.96, and IDSW 1905 (Fang, 2024). ACV replaces a one-step constant-velocity estimate with

$\rightarrow$6

and the paper reports that performance improves when $\rightarrow$7, stabilizes as $\rightarrow$8 increases, and is best at $\rightarrow$9 on MOT17-train (Fang, 2024).

The appearance-side change replaces the median element embedding of a tracklet with a confidence-aware update

$T = 0.5$0

where $T = 0.5$1 is adapted from detection confidence (Fang, 2024). In the corresponding ablation, Median App. gives HOTA 77.47, MOTA 87.35, IDF1 86.74, and AssA 77.58, while Dynamic App. gives HOTA 77.70, MOTA 87.30, IDF1 87.06, and AssA 77.96 (Fang, 2024). On MOT17-test, the full tracker reports HOTA 63.1, MOTA 76.9, and IDF1 78.2, with post-processing by AFlink + GSI increasing results to HOTA 64.1, MOTA 79.2, and IDF1 79.0 (Fang, 2024).

A plausible implication is that, even without explicit domain adaptation, the choice of geometric, motion, and appearance abstractions strongly conditions cross-scene robustness. The literature above repeatedly treats identity preservation as a representation problem as much as an optimization problem.

5. “MOT” in cold-atom physics and computational electromagnetics

Outside computer vision, the same abbreviation denotes unrelated technical objects. In cold-atom physics, “Demonstration of a MOT in a Sub-Millimeter Membrane Hole” defines a membrane MOT as a magneto-optical trap inside a sub-millimeter hole in a transparent membrane (Lee et al., 2020). The membrane is $T = 0.5$2, the main demonstration uses a $T = 0.5$3 hole, and six cooling beams intersect at the hole while the membrane splits the loading space into two hemispherical capture volumes (Lee et al., 2020). With $T = 0.5$4 on the D2 line at 852 nm and polarization-gradient cooling for about 1–2 ms, the paper reports cooling down to $T = 0.5$5, with lowest measured temperatures of $T = 0.5$6 for the unbridged $T = 0.5$7 hole and $T = 0.5$8 for the bridged $T = 0.5$9 hole (Lee et al., 2020). The atom number in the membrane MOT is about $T = 0.1$0 to $T = 0.1$1, and the $T = 0.1$2-diameter of the cloud is about $T = 0.1$3 for the $T = 0.1$4 hole; the device is presented as a loading stage for future evanescent-field optical traps in a photonic atom trap integrated platform (ATIP) (Lee et al., 2020). A frequent misconception is corrected explicitly in that paper: the membrane MOT is not the final trap, but a preload mechanism for suspended membrane waveguides and strong atom-light interaction (Lee et al., 2020).

In computational electromagnetics, MOT denotes the marching-on-in-time solution of the time-domain PMCHWT equation. “On the Late-Time Instability of MOT solution to the Time-Domain PMCHWT Equation” studies the late-time instability of this discretization and identifies the static solenoidal nullspace of the time-domain electric field integral operator as the primary cause (Le et al., 18 Sep 2025). The paper emphasizes that the instability mechanism is fundamentally different from that of the time-domain EFIE: PMCHWT is much more sensitive to numerical quadrature errors, and its spectral characteristics are strongly influenced by topology and surface smoothness (Le et al., 18 Sep 2025). On a sphere, for example, the reported eigenvalue shift is $T = 0.1$5 at $T = 0.1$6 and $T = 0.1$7 at $T = 0.1$8, whereas the TD-EFIE shift remains around $T = 0.1$9 and is largely round-off limited (Le et al., 18 Sep 2025). The paper also reports dramatic geometry dependence, with the star pyramid giving $N = 40$0 at $N = 40$1 and $N = 40$2 at $N = 40$3 (Le et al., 18 Sep 2025).

6. Domain specialists, routing, and validity-gated testing

The broader “domain-aware” literature adds two further motifs: specialization by domain and validation by domain. Meta-DMoE addresses multi-source unsupervised test-time adaptation by training a separate expert for each source domain, combining their outputs through a transformer-based aggregator, and distilling the result into a student updated on a small unlabeled target support set (Zhong et al., 2022). The paper’s central claim is that a single model trained across all sources is biased toward domain-invariant features and may suffer negative transfer, whereas separate experts preserve specialization and the aggregator is meta-trained to distill positive knowledge (Zhong et al., 2022). This notion of domain-aware specialization is conceptually adjacent to Domain-MoT, even though the task is not MOT.

MoDEM applies a related specialization principle to LLMs. It uses a DeBERTa-v3-large router with 304M parameters to assign each prompt to one of five domains—Math, Health, Science, Coding, or Other—and then activates only the corresponding expert model (Simonds et al., 2024). The router reports 97% average accuracy on held-out test data from its training datasets and 81.00% overall on manually mapped MMLU domains; the paper attributes the weaker “Other” accuracy to the fact that it is a broad catch-all category (Simonds et al., 2024). The system’s main claim is improved performance-to-cost ratio rather than universality: Medium MoDEM reports 87.7% on MMLU at an estimated \$N = 40$$40.88 for <a href="https://www.emergentmind.com/topics/compact-editor-model-llama-3-1-70b" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Llama 3.1 70B</a> and 88.6$$5.00 for Llama 3.1 405B (Simonds et al., 2024).

A different meaning of “domain-aware” appears in “Domain-Validity-Gated Metamorphic Testing of Scientific ML Surrogates,” where the relevant domain is the validity domain of a candidate metamorphic relation rather than a dataset or sensor modality (Li et al., 16 Jun 2026). A candidate relation is admitted only if it has a physical or software basis, its transformation preconditions hold, boundary-condition compatibility is preserved, and the verdict tolerance dominates the operator’s numerical floor (Li et al., 16 Jun 2026). The paper formalizes executable MR-cards and typed verdicts such as pass, fail, skip, out-of-relation-domain, numerical-tolerance issue, and inconclusive (Li et al., 16 Jun 2026). Its case studies sharply separate valid and invalid uses of symmetry or conservation: node permutation on MeshGraphNets cylinder flow holds to machine precision with relative $N = 40$5, mirror-y on the real asymmetric evaluation mesh is downgraded to an out-of-domain stress probe with median relative $N = 40$6, and incompressible continuity is rejected outright on a compressible airfoil task because density varies by a median factor $N = 40$7 (Li et al., 16 Jun 2026).

7. Formal domain theories and domain-aware communication

In formal semantics and concurrency theory, “domain” is not a dataset or sensor category but a mathematical or logical object. “Domain-Aware Session Types” extends the Curry–Howard interpretation of binary session types with modal worlds that indicate domains, plus an accessibility relation governing domain migration (Caires et al., 2019). Processes are typed not only by protocol but also by location, in judgments of the form

$N = 40$8

and the framework ensures session fidelity, global progress, termination, and that processes communicate only with accessible domains (Caires et al., 2019). The paper further introduces domain-aware multiparty session types, where global protocols can express arbitrarily nested sub-protocols via domain migration and are analyzed by reduction to the binary domain-aware framework (Caires et al., 2019).

“Mathematics of Domains” develops a different formal meaning of domain, centered on Scott topology, information systems, subdomains, retractions, valuations, and relaxed metrics (Bukatin, 2015). In the algebraic case, the paper proves the contravariance lemma

$N = 40$9

and shows that the collection of algebraic subdomains of a given domain forms an algebraic Scott domain (Bukatin, 2015). In its analytic half, it shows that the axiom $N = 20$0 is incompatible in general with Scott continuity of distance functions, motivating relaxed metrics and co-continuous valuations on Scott open sets (Bukatin, 2015). This suggests a far more abstract reading of Domain-MoT: domain sensitivity can be treated not merely as a transfer-learning nuisance, but as a structural property of information, topology, and communication.

Taken together, these literatures show that “Domain-MoT” is not a single standardized term but a family of domain-explicit problem formulations. In MOT proper, the dominant themes are cross-domain transfer, domain-specific sensing, and identity-preserving association under shifted conditions. In adjacent work, the same emphasis on domain specificity reappears as expert routing, validity gating, logical world structure, or topological semantics. The common thread is not a shared implementation, but the insistence that domain structure—whether dataset, sensor, physical validity region, microfabricated geometry, or logical world—must be represented explicitly rather than treated as background noise.