Paper deep dive
Template-Search Domain Adaptation via Multi-Stage Feature Alignment for Cross-Modal Object Tracking
Fereshteh Aghaee Meibodi, Amir Mehdi Soufi Enayati, Shadi Alijani, Homayoun Najjaran
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/1/2026, 8:45:36 AM
Summary
The paper introduces TSDA-Track, a template-search domain adaptation framework for cross-modal object tracking that reduces modality discrepancies between template and search frames. It proposes two feature alignment strategies: Pre-AFA (adversarial alignment before transformer interaction) and Enc-CFA (contrastive alignment after encoder interaction). Experiments on LasHeR, RGBT234, GTOT, and Anti-UAV-024 datasets demonstrate that TSDA-Track outperforms state-of-the-art trackers like ToMP-101, particularly in modality-switch scenarios, while maintaining a shared inference pipeline without modality-specific branches.
Entities (11)
Relation Signals (10)
Cross-Modal Object Tracking → addressedby → TSDA-Track
confidence 96% · This motivates the distinct setting of cross-modal object tracking... Accordingly, we introduce TSDA-Track
Enc-CFA → implements → Contrastive Feature Alignment
confidence 95% · Enc-CFA TSDA-Track applies contrastive alignment... Contrastive Feature Alignment (CFA) operates on jointly contextualized encoder outputs
Pre-AFA → implements → Adversarial Feature Alignment
confidence 95% · Pre-AFA TSDA-Track applies adversarial alignment... Adversarial Feature Alignment (AFA) suppresses modality-specific cues
TSDA-Track → uses → Pre-AFA
confidence 95% · Pre-AFA TSDA-Track applies adversarial alignment before transformer's template-search interaction
TSDA-Track → uses → Enc-CFA
confidence 95% · Enc-CFA TSDA-Track applies contrastive alignment to encoder representations after interaction
TSDA-Track → outperforms → ToMP-101
confidence 92% · Pre-AFA TSDA-Track achieves an SR/PR of 43.2/56.0, compared with 36.8/50.0 for ToMP-101 baseline.
TSDA-Track → evaluatedon → LasHeR
confidence 90% · Experiments on LasHeR... demonstrate improvements over representative state-of-the-art trackers.
TSDA-Track → evaluatedon →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Visual object tracking typically assumes that the initial template and subsequent search frames share the same sensing modality. In practice, sensor availability or operation may change over time, creating a substantial representation gap between template and search frames. Unlike conventional multi-modal tracking where paired modalities are simultaneously available, cross-modal tracking requires localization when template and search frames originate from different active modalities. Accordingly, we introduce TSDA-Track, a Template-Search Domain Adaptation framework to reduce modality discrepancy during training. We investigate two feature alignment strategies. Pre-AFA TSDA-Track applies adversarial alignment before transformer's template-search interaction to suppress modality-specific bias. Enc-CFA TSDA-Track applies contrastive alignment to encoder representations after interaction to strengthen target-level cross-modal correspondence. Both variants retain a shared inference pipeline without modality-specific branches. Experiments on LasHeR, and zero-shot evaluations on RGBT234 and GTOT under multiple cross-modal protocols demonstrate improvements over representative state-of-the-art trackers. For instance, under the modality-switch protocol on RGBT234, Pre-AFA TSDA-Track achieves an SR/PR of 43.2/56.0, compared with 36.8/50.0 for ToMP-101 baseline. In addition, a study on Anti-UAV-024 further verifies the applicability of TSDA-Track to aerial tracking. Our study highlights the effectiveness of feature alignment domain adaptation for cross-modal tracking.
Tags
Links
- Source: https://arxiv.org/abs/2609.38637v1
- Canonical: https://arxiv.org/abs/2609.38637v1
Trouble viewing inline? Open PDF directly →
Full Text
54,087 characters extracted from source content.
Expand or collapse full text
Template-Search Domain Adaptation via Multi-Stage Feature Alignment for Cross-Modal Object Tracking Fereshteh Aghaee Meibodi Affiliation: Department of Electrical and Computer Engineering Affiliation: University of Victoria Affiliation: 3800 Finnerty Road Affiliation: Victoria, BC, Canada Email: fereshtehaghaee@uvic.ca Amir Mehdi Soufi Enayati Affiliation: Department of Mechanical Engineering Affiliation: University of Victoria Affiliation: 3800 Finnerty Road Affiliation: Victoria, BC, Canada Email: amsoufi@uvic.ca Shadi Alijani Affiliation: Department of Mechanical Engineering Affiliation: University of Victoria Affiliation: 3800 Finnerty Road Affiliation: Victoria, BC, Canada Email: shadialijani@uvic.ca Homayoun Najjaran Affiliation: Department of Mechanical Engineering Affiliation: University of Victoria Affiliation: 3800 Finnerty Road Affiliation: Victoria, BC, Canada Email: najjaran@uvic.ca Abstract Visual object tracking typically assumes that the initial template and subsequent search frames share the same sensing modality. In practice, sensor availability or operation may change over time, creating a substantial representation gap between template and search frames. Unlike conventional multi-modal tracking where paired modalities are simultaneously available, cross-modal tracking requires localization when template and search frames originate from different active modalities. Accordingly, we introduce TSDA-Track, a Template-Search Domain Adaptation framework to reduce modality discrepancy during training. We investigate two feature alignment strategies. Pre-AFA TSDA-Track applies adversarial alignment before transformer’s template-search interaction to suppress modality-specific bias. Enc-CFA TSDA-Track applies contrastive alignment to encoder representations after interaction to strengthen target-level cross-modal correspondence. Both variants retain a shared inference pipeline without modality-specific branches. Experiments on LasHeR, and zero-shot evaluations on RGBT234 and GTOT under multiple cross-modal protocols demonstrate improvements over representative state-of-the-art trackers. For instance, under the modality-switch protocol on RGBT234, Pre-AFA TSDA-Track achieves an SR0.5/PR of 43.2/56.0, compared with 36.8/50.0 for ToMP-101 baseline. In addition, a study on Anti-UAV-024 further verifies the applicability of TSDA-Track to aerial tracking. Our study highlights the effectiveness of feature alignment domain adaptation for cross-modal tracking. A Preprint Figure 1: Comparison of (a) single-modal, (b) multi-modal RGB-T, and (c) proposed cross-modal tracking. Switches denote adaptive alignment strategy and interaction stage. 1 Introduction Single object tracking (SOT) in computer vision aims to localize an arbitrary target throughout a video sequence given only its initial state in the first frame. Over the past decade, single-modal trackers, shown in Figure 1(a), have achieved remarkable progress on standard benchmarks, with various designs ranging from discriminative-based methods and Siamese networks to recent transformer-based architectures. However, they generally assume that the template and search frames share the same visual modality. To enhance robustness, multi-modal object trackers, such as RGB-T trackers, fuse complementary information from different sensors, shown in Figure 1(b). Although effective, they typically require all modalities to be simultaneously available and synchronized at each time step. Thus, both single-modal and conventional multi-modal tracking assume a consistent template-search modality configuration. This limits their applicability in modality switch scenarios, where the template and search frames may originate from different domains due to sensor failure, occlusion of one sensor, low illumination, or temporal switching. The resulting modality mismatch introduces a template-search domain gap that can weaken feature interaction and degrade target localization. This motivates the distinct setting of cross-modal object tracking, which aims to accurately localize a target when the template and search frames may originate from different sensing domains, such as RGB and thermal imagery. Previous efforts have addressed this scenario through modality-aware mechanisms, including modality prediction Liu et al. (2025), adaptive fusion weighting Liu et al. (2024a); Li et al. (2022), and state-dependent routing Xu et al. (2025). These methods primarily select, weight, or route modality-dependent representations at inference time. Consequently, they do not explicitly optimize the template-search representation gap caused by sensing-domain mismatch, leaving feature interaction sensitive to the underlying distribution shift across domains. To address this issue, this paper proposes two variants of TSDA-Track, a template-search domain adaptation framework that reduces cross-modal discrepancy through feature alignment, illustrated in Figure 1(c). Rather than relying on modality prediction, adaptive fusion, or state-dependent routing, TSDA-Track explicitly aligns template and search representations across modality domains. Specifically, we investigate two alignment strategies. Pre-AFA TSDA-Track applies adversarial domain adaptation Ganin et al. (2016) that encourages modality-invariant representations before template-search encoder interaction. Enc-CFA TSDA-Track applies contrastive feature alignment Oord et al. (2018); Pang et al. (2021) which promotes target-level cross-modal consistency after encoder. Feature alignment improves tracking robustness in dynamic environments with severe modality shift without requiring explicit modality branches at inference. A summary of our main contributions is: • We formulate cross-modal tracking as a template-search domain mismatch perspective and evaluate it under fixed and temporally dynamic modality-switching protocols. • We propose TSDA-Track, a training-only domain-adaptation framework that reduces modality-induced template-search discrepancy through feature alignment. Since the alignment branches are removed after training, TSDA-Track retains a shared and modality-agnostic inference pipeline requiring no explicit modality estimation, adaptive fusion, modality-specific branches, or routing. • We develop two feature-alignment variants. Pre-AFA suppresses modality-specific bias via adversarial alignment, while Enc-CFA promotes target-level cross-modal consistency via contrastive alignment. Comprehensive ablations show that cross-modal performance depends on alignment objective, strength, and integration stage. • TSDA-Track shows better overall performance compared with representative trackers across multiple cross-modal protocols on RGB-T datasets, including LasHeR and zero-shot RGBT234/GTOT evaluations, and is also validated on Anti-UAV aerial tracking. 2 Related Work Single Object Tracking. SOT localizes an arbitrary target across video frames given its initial bounding-box. Existing SOT methods broadly fall into discriminative, Siamese-based, and transformer-based trackers Meibodi et al. (2025). Discriminative trackers Danelljan et al. (2019); Bhat et al. (2019) learn target-specific appearance models online, whereas Siamese-based trackers Yu et al. (2020); Liu et al. (2024b) formulate tracking as template-search matching using learned similarity functions. More recently, transformer-based trackers have advanced SOT through improved relation modeling. Early transformer-based trackers, such as TransT Chen et al. (2021) and STARK Yan et al. (2021), introduced attention mechanisms for robust feature fusion. ToMP Mayer et al. (2022) reformulated discriminative model prediction with a transformer encoder-decoder and was later extended to multi-object tracking in TaMOs Mayer et al. (2024). OSTrack Ye et al. (2022a) and MixFormer Cui et al. (2024) apply one-stream frameworks that jointly process template and search tokens. Recent advances include autoregressive sequence prediction in ARTrack Wei et al. (2023) and SeqTrack Chen et al. (2023), spatio-temporal sequence modeling in ODTrack Zheng et al. (2024) and MCITrack Kang et al. (2025), and temporally-aware representation modeling in ROMTrack Cai et al. (2023). However, most SOT trackers assume visually consistent, same-modality template and search frames, making them vulnerable to modality switches and template-search discrepancies. Multi-Modal and Cross-Modal Object Tracking. To improve robustness under illumination variation, occlusion, and adverse imaging conditions, multi-modal tracking fuses complementary cues from different sensing modalities, addressing single-modal limitations. In RGB-T tracking, APFNet Xiao et al. (2022) explores attribute-aware fusion and MTNet Hou et al. (2023) employs transformers to model global inter-modal relations and better exploit complementary cues. Recent RGB-X trackers, such as ProTrack Yang et al. (2022), ViPT Zhu et al. (2023), and SDSTrack Hou et al. (2024), introduce prompt-based or unified adaptation mechanisms to improve multi-modal integration. However, these methods assume simultaneously available paired modalities and are not designed for modality-switching scenarios with only one active modality. Cross-modal object tracking relaxes this assumption by allowing modality-switching scenarios or template-search frames from different sensing domains. ProtoTrack Liu et al. (2025) updates modality-specific prototypes through modality classification. MArMOT and MAFNet Li et al. (2022); Liu et al. (2024a) learn adaptive weights for RGB-NIR branches, and SwiTrack Xu et al. (2025) applies tri-state routing. These cross-modal approaches rely on modality-aware inference mechanisms through modality prediction, adaptive weighting, or state-dependent routing. In contrast, the proposed Pre-AFA and Enc-CFA TSDA-Track directly reduce the template-search domain gap through training-only feature alignment, while retaining a shared inference pipeline without modality-specific branches or routing. Domain Adaptation for Tracking. Domain Adaptation reduces distribution shift between training and target domains. In visual tracking, it has mainly been explored to adapt models across environments or sensing domains. UDAT Ye et al. (2022b) aligns daytime and nighttime features via adversarial day-night discrimination, while subsequent work improves nighttime adaptation through transformer-based fusion Wei et al. (2024), temporal context alignment Fu et al. (2024a), or SAM-assisted target-domain sample mining Fu et al. (2024b). For thermal tracking, PDAT Li et al. (2024) transfers RGB knowledge to TIR by progressively aligning global and subdomain feature distributions, while EHDA Li et al. (2025) performs hierarchical RGB-TIR adaptation at style and semantic levels. Trans-DAT Wu et al. (2024) further applies gradient reversal to reduce discrepancies among hyperspectral sensor domains. Unlike these methods that adapt trackers at the dataset or model level, the proposed TSDA-Track targets the within-sequence domain gap between cross-modal template and search frames. 3 Proposed Method Figure 2: Overview of the proposed TSDA-Track for cross-modal tracking. AFA and CFA denote selectable offline-training alignment branches. Pre-AFA aligns pooled template and search features before the encoder via adversarial learning, whereas Enc-CFA aligns pooled encoder representations through contrastive learning. Figure 2 illustrates TSDA-Track, built on the transformer-based ToMP Mayer et al. (2022) tracker. Given target-annotated templates =tii=1NtT=\t_i\_i=1^N_t and a search frame s, a shared backbone extracts template and search features, respectively augmented with target-state and learnable test-frame encodings. A transformer encoder jointly processes these representations, and the decoder predicts target-specific model weights to estimate the target state in s. TSDA-Track addresses modality-mismatched tracking, where the template and search frames may originate from different sensing domains, including RGB and thermal imagery. Instead of modality-specific branches or routing, TSDA-Track applies training-only template-search feature-alignment to reduce modality-induced discrepancy while preserving the discriminative information for accurate tracking. 3.1 Cross-Modal Tracking Setting Unlike standard SOT frameworks that assume same-modality template and search frames, cross-modal tracking allows them to originate from different sensing domains, such as RGB and thermal imagery, with distinct appearance statistics, contrast, and background cues. The resulting distribution shift can impair template-search interaction in the transformer and degrade target-state estimation. This motivates the feature-alignment objectives introduced below. 3.2 Domain Adaptation for Cross-Modal Tracking To reduce template-search modality discrepancy while preserving target-discriminative features, TSDA-Track investigates two alternative training-only alignment strategies with distinct objectives and operating stages. Adversarial Feature Alignment (AFA) suppresses modality-specific cues before joint template-search encoding, promoting modality-invariant representations. Contrastive Feature Alignment (CFA) operates on jointly contextualized encoder outputs to strengthen target-conditioned cross-modal consistency. Each auxiliary alignment loss is jointly optimized with the tracking classification and regression losses and removed at inference. TSDA-Track thus retains the standard shared template-search pipeline without modality prediction, modality-specific branches, or routing decisions. 3.2.1 Adversarial Feature Alignment Adversarial domain adaptation is applied before the transformer encoder jointly processes template and search features. This provides a more compatible representation space for relation modeling because features at this stage may retain modality-specific appearance statistics that bias subsequent template-search interaction. Let Ft,ipreF^pre_t,i and FspreF^pre_s denote the pre-encoder feature maps of the i-th template and search frame, respectively. Since the template frames share the same modality during training, their pooled descriptors are aggregated as ftpre=1Nt∑i=1NtGAP(Ft,ipre),fspre=GAP(Fspre).f^pre_t= 1N_t _i=1^N_tGAP (F^pre_t,i ),\ \ f^pre_s=GAP (F^pre_s ). (1) where GAP(⋅)GAP(·) denotes global average pooling. We apply adversarial alignment to these pooled descriptors rather than spatial feature maps because directly modifying spatial features may suppress the location-sensitive cues required for target localization in the tracking pipeline. Let mt,ms∈0,1m_t,m_s∈0,1 denote the modality labels of the template and search descriptors, where 00 and 11 correspond to RGB and thermal imagery, respectively. A shared pre-encoder domain discriminator DpreD_pre predicts the modality of each descriptor. To adversarially train the shared feature extractor, a gradient-reversal layer (GRL) Ganin and Lempitsky (2015) is inserted before the discriminator. The GRL acts as an identity mapping during the forward pass while multiplying the gradient propagated to the feature extractor by −γ-γ during backpropagation, resulting in −γ∇fprex,x∈t,s-γ∇f^pre_x,x∈t,s. Here, γ controls the strength of the adversarial signal. Thus, the discriminator distinguishes modalities while the shared feature extractor learns modality-invariant representations. The domain-classification loss is defined as ℒadv=12∑x∈t,sℓBCE(Dpre(fxpre),m(x)).L_adv= 12 _x∈t,s _BCE (D_pre(f^pre_x),m(x) ). (2) where ℓBCE _BCE denotes binary cross-entropy loss. The discriminator is optimized to identify the sensing modality, whereas the reversed gradient encourages the shared feature extractor to deceive it by making modality prediction difficult. This encourages modality-invariant representations and reduces modality-specific discrepancy before the transformer encoder jointly processes template and search information. 3.2.2 Contrastive Feature Alignment While Pre-AFA reduces modality discrepancy before template-search interaction, CFA regularizes jointly contextualized representations produced by the transformer encoder. At this stage, the encoder has already modeled interactions between target-conditioned template and current search representations. The resulting features therefore contain information directly relevant to subsequent target-model prediction. CFA exploits this information to encourage cross-modal semantic consistency between corresponding template-search pairs while preserving discrimination among different tracking tuples. Let Ft,iencF^enc_t,i and FsencF^enc_s denote the post-encoder representations of the i-th template and search frame, respectively. Since template frames share the same modality during training, their global descriptors are aggregated as ftenc=1Nt∑i=1NtGAP(Ft,ienc),fsenc=GAP(Fsenc).f^enc_t= 1N_t _i=1^N_tGAP (F^enc_t,i ),\ \ f^enc_s=GAP (F^enc_s ). (3) We perform contrastive alignment on pooled encoder descriptors rather than enforcing token-wise spatial correspondence because RGB and thermal features can exhibit different local appearance and contrast responses. A shared projection head P(⋅)P(·) maps the pooled template and search descriptors into a shared embedding space, yielding projected template and search embeddings zt=P(ftenc)z_t=P(f^enc_t) and zs=P(fsenc)z_s=P(f^enc_s), respectively. For a mini-batch of B tracking tuples, the ℓ2 _2-normalized embeddings construct a batch-wise template-search similarity matrix: z¯ti=zti‖zti‖2,z¯sj=zsj‖zsj‖2,i,j=1,…,B. z_t^\,i= z_t^\,i \|z_t^\,i \|_2, z_s^\,j= z_s^\,j \|z_s^\,j \|_2, i,j=1,…,B. (4) Aij=sim(z¯ti,z¯sj)τ=z¯ti⊤z¯sjτ,A=[Aij]i,j=1B.A_ij= sim ( z_t^\,i, z_s^\,j )τ= z_t^\,i z_s^\,jτ, A=[A_ij]_i,j=1^B. (5) where sim(⋅,⋅)sim(·,·) denotes cosine similarity and τ is a temperature parameter. Since the embeddings are normalized, their dot product equals cosine similarity. The diagonal entries AiiA_i represent template and search embeddings from the same tracking tuple and are treated as positive pairs, whereas off-diagonal entries represent non-corresponding tuples in the mini-batch and serve as negatives. The template-to-search contrastive objective is defined as ℒt→s=−1B∑i=1Blogexp(Aii)∑j=1Bexp(Aij).L_t→ s=- 1B _i=1^B (A_i) _j=1^B (A_ij). (6) Similarly, the search-to-template objective is computed by reversing the matching direction: ℒs→t=−1B∑i=1Blogexp(Aii)∑j=1Bexp(Aji).L_s→ t=- 1B _i=1^B (A_i) _j=1^B (A_ji). (7) The final encoder-level contrastive loss is computed symmetrically as ℒctrenc=12(ℒt→s+ℒs→t).L_ctr^enc= 12 (L_t→ s+L_s→ t ). (8) This bidirectional objective keeps each template representation close to its corresponding search-frame representation after encoder interaction while separating non-corresponding mini-batch samples. Jointly optimized with tracking losses, CFA promotes cross-modal consistency in encoder representations without discarding the discriminative information needed for target classification and bounding-box regression. 3.3 Offline Training Each training sample contains two target-annotated templates =T1,T2T=\T_1,T_2\ and one search frame s. The templates are sampled from the same modality, m(T1)=m(T2)m(T_1)=m(T_2), to stabilize model prediction during training. The search frame is sampled from either the same or opposite modality with pmismatch=0.5p_mismatch=0.5, yielding balanced same-modality and cross-modal template-search pairs. For cross-modal samples, RGB-to-thermal and thermal-to-RGB template-search pairs are sampled uniformly. Training objective. TSDA-Track augments the standard ToMP tracking objective with an auxiliary template-search alignment loss. The classification loss ℒclsL_cls supervises target score map using a Gaussian label Bhat et al. (2019), while ℒGIoUL_GIoU supervises dense bounding-box predictions Mayer et al. (2022). The objectives for Pre-AFA and Enc-CFA are ℒAFA=λclsℒcls+λregℒGIoU+λAFAℒadvpre,L_AFA= _clsL_cls+ _regL_GIoU+ _AFAL_adv^pre, (9) ℒCFA=λclsℒcls+λregℒGIoU+λCFAℒctrpost,L_CFA= _clsL_cls+ _regL_GIoU+ _CFAL_ctr^post, (10) where ℒadvpreL_adv^pre is the pre-encoder modality-classification loss and ℒctrpostL_ctr^post is the encoder-level contrastive loss. The weights λAFA _AFA and λCFA _CFA control the contribution of alignment losses. 3.4 Online Tracking During online tracking, AFA and CFA branches are disabled, so TSDA-Track follows a shared and modality-agnostic ToMP inference pipeline without modality labels and modality-specific branches. Unlike approaches that explicitly identify or fuse modalities using inference modality classifiers or learned routing weights, TSDA-Track requires no such inference-time modules. This limits error propagation from inaccurate modality estimates, fusion weights, or routing and reduces sensitivity to whether the test modalities match those used to train inference-time modules. During online template memory update, the annotated initial template is retained while the remaining slots are updated with confident predictions. Thus, unlike offline training, the online memory may naturally contain RGB, thermal, or mixed-modality frames due to the input stream and update history. 4 Experiments 4.1 Implementation Details We implement TSDA-Track by extending ToMP Mayer et al. (2022). The transformer contains six encoder and six decoder layers. All variants are implemented in Python with PyTorch, Training is conducted on a single NVIDIA H100 GPU, and evaluation is performed on an NVIDIA TITAN RTX GPU. Training and Inference Details. We fine-tune all models from a pretrained ToMP-101 checkpoint on the LasHeR Li et al. (2021) training split for 3030 epochs using 4040k subsequences and a batch size of 88. Template and search crops are resized to 288×288288× 288, yielding 18×1818× 18 feature maps. The target filter size and search-area factor are set to 33 and 5.05.0. We use AdamW Loshchilov and Hutter (2017) with weight decay 10−410^-4, learning rates of 10−410^-4 for the tracking head and alignment modules and 10−510^-5 for ResNet layers, reduced by a factor of 0.20.2 after 2020 epochs. The backbone is unfrozen to better align the representation mismatch through alignments. Auxiliary alignment losses are enabled after five epochs to stabilize training. We set λcls=100 _cls=100 and λreg=5 _reg=5. For Pre-AFA, λAFA=0.5 _AFA=0.5 and the GRL coefficient γ is 0.10.1. For Enc-CFA, λCFA=0.1 _CFA=0.1 and τ is 0.10.1. In the joint variant, λAFA _AFA is 0.30.3 and λCFA _CFA is 0.10.1 with all other settings unchanged. Unlike ToMP, independent target-centered position and scale jitter are applied to both template and search crops to improve robustness under modality mismatch. All TSDA-Track variants use ToMP inference settings. CM-ToMP-101 and CM-MAFNet denote ToMP-101 and MAFNet, respectively, fine-tuned with cross-modal template–search pairs but without the proposed explicit feature-alignment modules. Experimental Setup We design three novel template-search modality-discrepancy protocols to assess robustness to fixed and dynamic modality shifts. RGB→ initializes the tracker with an RGB template and thermal search frames, whereas T→ reverses this setting. In Switch mode, the template is RGB and the search modality alternates between thermal and RGB every 200200 frames. We report success rate at IoU 0.50.5 (SR0.5) and precision rate at a 2020-pixel center-error threshold (PR) (see more SR/PR plots in suppl. material). 4.2 Ablation Study and Analysis We conduct an ablation study of TSDA-Track to examine how the alignment objective, its stage-specific integration, and alignment strength affect cross-modal robustness. Table 1: Zero-shot ablation on cross-modal training (CM), alignment strategy and stage on RGBT234 Li et al. (2019). Adv./Con. FA denote adversarial/contrastive alignment. Pre and Post refer to pre- and post-encoder alignment. The top two results are bolded and underlined, respectively. Method CM Train. Adv. FA Con. FA RGB→ T→ Switch SR0.5 PR SR0.5 PR SR0.5 PR ToMP-101 −- −- −- 31.7 44.9 33.2 43.1 36.8 50.0 CM-ToMP-101 ✓ −- −- 34.8 47.8 36.4 47.3 35.9 47.6 TSDA-Track variants: Pre-AFA ✓ Pre −- 40.5 53.9 39.9 52.4 43.2 56.0 Pre-CFA ✓ −- Pre 37.0 48.1 35.9 46.1 38.6 49.4 Enc-AFA ✓ Post −- 36.6 50.6 37.5 50.3 38.2 51.7 Enc-CFA ✓ −- Post 38.6 52.0 38.6 51.9 41.5 54.4 Pre-AFA + Enc-CFA ✓ Pre Post 39.7 50.7 42.6 53.2 42.6 53.6 4.2.1 Effect of Alignment Strategy and Stage Table 1 evaluates the effects of cross-modal training, alignment strategy, and alignment stage. Compared with ToMP-101, CM-ToMP-101 improves the fixed RGB→ and T→ settings but degrades Switch performance, showing that cross-modal training alone is insufficient for repeated modality changes. Explicit feature alignment improves robustness across protocols. Pre-AFA achieves the best RGB→ and Switch results, 40.5/53.940.5/53.9 and 43.2/56.043.2/56.0, respectively. Its advantage over Enc-AFA indicates that adversarial alignment is most effective before transformer interaction, where modality-specific appearance bias can be suppressed before template-search relation modeling in the encoder. In contrast, CFA is more effective after encoder interaction. Early features remain strongly influenced by modality-specific appearance and local contrast, making Pre-CFA less reliable for target-level alignment. Enc-CFA instead operates on contextualized template-search representations after encoder, where positive pairs better represent target-level correspondence and negatives provide more meaningful discriminative supervision. Pre-AFA remains stronger because it prevents modality bias from entering transformer interaction, whereas Enc-CFA cannot fully recover relation modeling already affected by modality mismatch. Combining the strongest individual configurations, Pre-AFA+Enc-CFA achieves the best T→ result (42.6/53.242.6/53.2) but does not consistently improve RGB→ or Switch. Thus, the objectives are not universally additive. Joint optimization may over-constrain the shared representation and attenuate target-discriminative cues required for accurate localization during tracking. 4.2.2 Effect of Alignment Strength Table 2 evaluates alignment strength on TSDA-Track variants. In Pre-AFA, the GRL coefficient scales the adversarial gradient on the shared feature extractor. γ=0.1γ=0.1 consistently outperforms γ=1.0γ=1.0, indicating that excessive adversarial alignment can suppress target-discriminative features needed for tracking. In Enc-CFA, a lower temperature τ sharpens the similarity distribution and increases contrastive pressure. τ=0.1τ=0.1 performs best across protocols, suggesting that moderate contrastive alignment better accommodates modality and appearance variation. Table 2: Zero-shot evaluation of the effect of Pre-AFA and Enc-CFA alignment strengths on RGBT234 Li et al. (2019). Method Parameter Strength RGB→ T→ Switch SR0.5 PR SR0.5 PR SR0.5 PR Pre-AFA TSDA-Track γ 0.10.1 40.5 53.9 39.9 52.4 43.2 56.0 1 38.5 49.5 39.7 49.7 40.7 51.5 Enc-CFA TSDA-Track τ 0.05 36.8 49.4 37.2 47.4 39.1 50.7 0.1 38.6 52.0 38.6 51.9 41.5 54.4 4.2.3 Feature-Level Analysis Figure 3 visualizes normalized pre-encoder representations of RGB templates and thermal searches on RGBT234. ToMP-101 exhibits clear modality separation, indicating a substantial pre-interaction RGB-thermal gap. CM-ToMP-101 increases overlap through cross-modal training but retains modality-specific structure. Pre-AFA TSDA-Track yields the strongest mixing by directly suppressing pre-encoder modality cues before transformer interaction. Enc-CFA also reduces the gap, but the distributions remain more structured as it promotes post-encoder target correspondence rather than pre-encoder modality invariance. To complement the PCA visualization, Table 3 quantifies identity-preserving cross-modal alignment at the pre- and post-encoder levels. For each RGB-template/thermal-search pair i, di,l+d_i,l^+ denotes the cosine distance between matched features from the same sequence, and di,l−d_i,l^- denotes the mean distance to unmatched thermal-search features from other sequences in the batch. We define Rloc=mediani(di,l+di,l−+ϵ),loc∈pre,post,R_loc=median_i ( d_i,l^+d_i,l^-+ε ), loc∈\pre,post\, (11) where ϵε is a small constant. Lower RlocR_loc indicates stronger same-object RGB-thermal alignment relative to different-object pairs. CM-ToMP-101 reduces both ratios compared with ToMP-101, confirming the benefit of cross-modal training. Pre-AFA achieves the lowest pre-encoder ratio (0.4350.435), consistent with its direct pre-interaction alignment objective, whereas Enc-CFA achieves the lowest post-encoder ratio (0.5940.594), matching its encoder-level contrastive supervision. These results quantitatively support that each alignment strategy is most effective at its intended feature stage. Table 3: Positive/negative cross-modal distance ratio on RGBT234 Li et al. (2019). Lower values indicate stronger same-object alignment relative to different-object pairs. Method Pre-encoder Rpre↓R_pre Post-encoder Rpost↓R_post Baseline:ToMP-101 Mayer et al. (2022) 0.730 0.862 CM-ToMP-101 0.496 0.625 Pre-AFA TSDA-Track 0.435 0.625 Enc-CFA TSDA-Track 0.474 0.594 Figure 3: PCA visualization of normalized pre-encoder representations on RGBT234 Li et al. (2019). 4.3 Comparison with Representative Trackers Table 4: Cross-modal evaluation against representative single-object trackers on RGBT234 Li et al. (2019), LasHeR Li et al. (2021), and GTOT Li et al. (2016) under RGB→ , T→ , and Switch protocols. RGBT234 and GTOT results are zero-shot. Bold, underlined, and italic values indicate the first-, second-, and third-best results, respectively. Methods RGBT234 Li et al. (2019) LasHeRLi et al. (2021) GTOT Li et al. (2016) RGB→ T→ Switch RGB→ T→ Switch RGB→ T→ Switch SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR SR0.5 PR ATOM Danelljan et al. (2019) 27.7 37.6 29.5 39.7 32.0 41.2 14.7 18.7 14.2 19.5 17.7 20.7 23.9 39.4 38.8 52.3 27.9 39.6 DiMP-50 Bhat et al. (2019) 24.9 32.8 32.9 44.0 33.3 41.3 15.3 17.3 18.4 22.6 19.6 22.4 27.3 38.7 30.0 45.0 27.4 40.0 KeepTrack Mayer et al. (2021) 26.0 34.2 29.2 36.8 36.7 44.8 16.6 18.8 18.0 20.4 21.6 23.9 28.1 41.2 25.7 37.0 29.1 41.6 STARK-ST Yan et al. (2021) 19.7 24.6 21.6 26.7 30.6 35.8 11.3 12.8 16.7 18.0 18.2 20.0 31.6 44.7 26.0 33.4 34.1 46.8 ToMP-101 Mayer et al. (2022) 31.7 44.9 33.2 43.1 36.8 50.0 21.6 25.7 19.4 24.3 25.2 29.3 40.7 60.4 37.7 55.6 39.9 59.9 OSTrack Ye et al. (2022a) 19.1 25.3 29.7 36.6 30.1 36.0 10.2 11.8 21.2 23.2 18.0 19.8 30.0 37.8 32.0 43.9 32.9 40.8 MixFormer-l Cui et al. (2024) 25.6 31.9 29.4 36.3 33.7 40.1 18.4 20.3 23.4 26.9 24.4 26.5 37.2 55.2 38.6 47.9 38.9 56.8 ODTrack-b Zheng et al. (2024) 25.4 42.6 44.0 55.7 33.4 47.5 14.9 19.7 30.4 35.5 19.4 23.6 37.8 54.4 40.9 52.2 39.6 56.4 SeqTrack-l384 Chen et al. (2023) 27.4 37.0 32.1 40.4 36.8 45.1 19.4 22.4 24.1 28.8 26.3 29.7 33.3 54.7 32.1 47.6 34.3 55.2 MCITrack-l384 Kang et al. (2025) 22.6 27.8 25.9 31.1 34.3 39.3 12.7 13.4 23.4 25.3 21.0 21.8 32.4 42.6 24.5 32.1 35.4 45.8 MAFNet Liu et al. (2024a) 20.8 36.8 25.2 45.7 24.2 41.3 13.1 20.3 14.6 22.5 14.3 21.7 25.3 47.4 26.3 60.8 25.2 47.7 Cross-modal trained: CM-MAFNet 32.0 54.7 30.4 57.2 34.9 58.3 21.8 33.6 23.4 36.8 22.5 35.6 28.7 52.7 37.3 72.0 31.0 55.3 CM-ToMP-101 34.8 47.8 36.4 47.3 35.9 47.6 26.3 30.3 29.6 32.4 27.3 31.2 47.3 62.8 39.3 56.4 46.3 62.0 Enc-CFA TSDA-Track 38.6 52.0 38.6 51.9 41.5 54.4 28.4 32.8 28.8 32.4 28.3 32.9 49.4 65.3 53.7 67.6 49.6 65.2 Pre-AFA TSDA-Track 40.5 53.9 39.9 52.4 43.2 56.0 36.6 38.9 37.7 40.0 37.3 39.8 45.7 61.3 52.9 70.0 45.8 61.8 Table 4 compares TSDA-Track with representative SOT trackers and the available cross-modal MAFNet Liu et al. (2024a) using their released checkpoints under the proposed RGB-T template-search modality-discrepancy protocols. Note that comparison with other cross-modal trackers is infeasible because of unavailable implementations or models. MAFNet’s RGB-NIR-trained modality-specific branches and its adaptive fusion weights transfer poorly to RGB-thermal, causing it to underperform several single-modal trackers. Therefore, cross-modal fusion weights learned for one sensor pair may not generalize to another without adaptation, as the improvement of CM-MAFNet over MAFNet further confirms the importance of sensor-pair adaptation. In addition, strong single-modal tracking does not necessarily transfer to cross-modal settings, and the CM-ToMP-101 and CM-MAFNet results further indicate that cross-modal training alone is insufficient for stable performance under modality discrepancy. RGBT234 and GTOT are evaluated zero-shot, without target-dataset training or fine-tuning. RGBT234 Li et al. (2019). Pre-AFA TSDA-Track achieves the highest zero-shot SR performance in RGB→ and Switch with 40.540.5 and 43.243.2, respectively. This indicates that suppressing modality bias before transformer interaction is beneficial where modality discrepancy is coupled with diverse target appearance and background changes. Compared with MAFNet, CM-MAFNet improves after RGB-T cross-modal fine-tuning and achieves high PR across protocols. However, its substantial lower SR indicates less consistent bounding-box overlap. ODTrack achieves the highest T→ SR of 44.044.0 benefiting from richer RGB search stream, but its lower RGB→ and Switch scores show that this advantage is direction-specific rather than consistent cross-modal robustness. Qualitative comparisons in Figure 4 support the quantitative findings that explicit template-search feature alignment improves tracking across modality transitions. Figure 4: Qualitative comparison on RGBT234 Li et al. (2019) under RGB→T, T→RGB, and Switch protocols. LasHeR Li et al. (2021). Pre-AFA TSDA-Track achieves the best performance across all protocols on LasHeR. Its consistent advantage under modality discrepancies indicates that pre-encoder alignment effectively suppresses modality bias before template-search interaction, yielding more robust target modeling where cross-modal discrepancy is coupled with LasHeR’s substantial appearance variation. GTOT Li et al. (2016). In zero-shot evaluation on GTOT, Enc-CFA TSDA-Track achieves the best SR0.5 under all three protocols. CM-MAFNet obtains the highest T→ PR of 72.072.0 but a substantially lower SR of 37.337.3, compared with 53.753.7 for Enc-CFA. GTOT trend differs from RGBT234 and LasHeR, where Pre-AFA is stronger. GTOT differs from the training data in acquisition conditions and exhibits a more diverse modality distribution, as it also includes grayscale input. This shift may limit Pre-AFA transferability, as its pre-encoder alignment can remain sensitive to input-domain statistics. In contrast, Enc-CFA aligns target-conditioned template-search representations after encoder interaction, with less explicit reliance on modality-specific low-level statistics. This may support stronger zero-shot transfer. 4.3.1 Attribute-Based Performance on RGBT234. To examine robustness under different tracking challenges, we evaluate the methods using the official sequence-level attributes of RGBT234 Li et al. (2019). For each attribute, SR0.5SR_0.5 and PR are averaged over all sequences based on the corresponding annotations. As shown in Figure 5, cross-modal training alone provides inconsistent gains over ToMP-101, confirming that exposure to mismatched template-search modalities does not sufficiently address their representation gap. Pre-AFA achieves the strongest overall attribute performance, with clear improvements under several challenging conditions, including occlusion, illumination variation, thermal crossover, deformation, scale change, motion blur, and camera motion. This supports the benefit of suppressing modality-dependent appearance bias before template-search interaction. Enc-CFA also improves multiple attributes by strengthening target-level correspondence after encoder interaction, although its gains on RGBT234 are less consistent than those of Pre-AFA. Performance remains comparatively limited under low-resolution and background-clutter, indicating that feature alignment alone cannot fully resolve information loss and target-distractor ambiguity. Figure 5: Attribute-based (SR0.5, PR) comparison on RGBT234 Li et al. (2019) under the RGB→ , T→ , and Switch protocols. 4.3.2 Success and Precision Plots. Figure 6 presents the full success and precision curves on RGBT234 Li et al. (2019) under RGB→ , T→ , and Switch cross-modality protocols. Pre-AFA TSDA-Track shows the most consistent overall robustness, with particularly promising success rate curves under RGB→ and Switch. Enc-CFA also consistently improves over ToMP-101 and CM-ToMP-101. ODTrack Zheng et al. (2024) achieves strong performance in T→ with the highest SR0.5 in this protocol, but this advantage is direction-specific, as its substantially weaker RGB→ and Switch performance indicates limited consistency across configurations. MAFNet Liu et al. (2024a) performs poorly under direct transfer from RGB-NIR to RGB-T. CM-MAFNet benefits from RGB-T cross-modal training and achieves high precision across all three protocols, but its SR remains substantially lower than that of the TSDA-Track variants indicating that improved target-center localization does not translate into equally reliable bounding-box overlap and scale estimation. Overall, the curves show that explicit feature alignment provides more consistent cross-modal robustness across protocols than either direction-specific gains or cross-modal training alone. (a) Success plots under the RGB→ , T→ , and Switch settings. (b) Precision plots under the RGB→ , T→ , and Switch settings. Figure 6: Tracking performance on RGBT234 Li et al. (2019) under the RGB→ , T→ , and Switch evaluation protocols. 4.3.3 Aerial Cross-Modal Tracking. We conduct a study on aerial RGB-T Anti-UAV-024 Jiang et al. (2021). Since its RGB and thermal videos have drastic spatial and resolution mismatch, we align the dataset using thermal frames as the reference and apply affine normalization preprocessing on RGB frames to reduce registration and scale mismatch. Models are trained on the training split and evaluated on the test split under the cross-modal protocols. Table 5 shows that Pre-AFA performs best under all protocols. Relative to CM-ToMP-101, it improves SR0.5/PR by 6.5/6.66.5/6.6, 5.2/4.95.2/4.9, and 4.4/4.14.4/4.1 points for RGB→ , T→ , and Switch, respectively. Enc-CFA consistently ranks second, while Pre-AFA+Enc-CFA remains inferior, indicating that they are not reliably additive as joint alignment may suppress useful tracking information. Qualitative results are shown in Figure 7. Figure 7: Visual comparison on Anti-UAV-024 Jiang et al. (2021) test split under RGB→ , T→ , and Switch mode. 5 Limitations TSDA-Track improves cross-modal robustness through feature alignment but does not explicitly address geometric mismatch, which may limit bounding-box regression under pixel-level misalignment, scale or aspect-ratio variation, and clutter. Future work will explore joint feature geometry-aware alignment, develop large-scale benchmarks covering broader modality transitions, and extend TSDA-Track to additional modality pairs such as RGB-NIR, RGB-depth, and event data. Moreover, TSDA-Track is architecture-dependent rather than backbone-agnostic as its alignment modules are built on ToMP with explicit access to pre- and post-interaction stages. Extending it to one-stream trackers such as OSTrack and MixFormer requires alignment reformulation since template and search tokens interact from early stages. Table 5: Cross-modal tracking performance on Anti-UAV-024 Jiang et al. (2021) dataset. Best and second-best results are shown in bold and underline, respectively. Method RGB→ T→ Switch SR0.5 PR SR0.5 PR SR0.5 PR ToMP-101 Mayer et al. (2022) 61.9 63.9 56.1 67.1 69.5 71.9 CM-ToMP-101 69.7 71.9 77.8 79.8 75.2 77.9 Enc-CFA TSDA-Track 74.3 78.1 81.5 84.2 77.4 81.2 Pre-AFA TSDA-Track 76.2 78.5 83.0 84.7 79.6 82.0 Pre-AFA + Enc-CFA TSDA-Track 72.0 75.7 79.5 83.5 76.2 79.8 6 Conclusions This paper introduced TSDA-Track, a template-search domain adaptation framework for cross-modal tracking with two training-only alignment strategies. Pre-AFA suppresses modality-specific bias through adversarial alignment before transformer template-search interaction at encoder, while Enc-CFA strengthens target-level cross-modal correspondence through post-encoder contrastive alignment. Zero-shot results on RGBT234 and GTOT, together with evaluations on LasHeR and Anti-UAV-024, show that explicit alignment outperforms representative single-modal and prior cross-modal trackers across cross-modal switching protocols while retaining a shared inference pipeline without modality-specific routing. CM-ToMP-101 and CM-MAFNet further show that cross-modal training alone is insufficient for stable performance under modality discrepancy. Pre-AFA is more consistent under broader appearance variation, while Enc-CFA shows stronger transfer under shifts in acquisition conditions and modality-specific statistics, benefiting from its target-level semantic alignment. This highlights the importance of alignment objective and its integration stage in cross-modal tracking. References Liu et al. [2025] Lei Liu, Chenglong Li, Futian Wang, Longfeng Shen, and Jin Tang. Prototype-based cross-modal object tracking. Information Fusion, 118:102941, 2025. Liu et al. [2024a] Lei Liu, Mengya Zhang, Cheng Li, Chenglong Li, and Jin Tang. Cross-modal object tracking via modality-aware fusion network and a large-scale dataset. IEEE Transactions on Neural Networks and Learning Systems, 36(4):6981–6994, 2024a. Li et al. [2022] Chenglong Li, Tianhao Zhu, Lei Liu, Xiaonan Si, Zilin Fan, and Sulan Zhai. Cross-modal object tracking: Modality-aware representations and a unified benchmark. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1289–1296, 2022. Xu et al. [2025] Boyue Xu, Ruichao Hou, Tongwei Ren, Dongming Zhou, Gangshan Wu, and Jinde Cao. Switrack: Tri-state switch for cross-modal object tracking. arXiv preprint arXiv:2511.16227, 2025. Ganin et al. [2016] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016. Oord et al. [2018] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. Pang et al. [2021] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 164–173, 2021. doi:10.1109/CVPR46437.2021.00023. Meibodi et al. [2025] Fereshteh Aghaee Meibodi, Shadi Alijani, and Homayoun Najjaran. A deep dive into generic object tracking: A survey. arXiv preprint arXiv:2507.23251, 2025. Danelljan et al. [2019] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. Bhat et al. [2019] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. Yu et al. [2020] Yuechen Yu, Yilei Xiong, Weilin Huang, and Matthew R. Scott. Deformable siamese attention networks for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. Liu et al. [2024b] Jing Liu, Han Wang, Chao Ma, Yuting Su, and Xiaokang Yang. Siamdmu: Siamese dual mask update network for visual object tracking. IEEE Transactions on Emerging Topics in Computational Intelligence, 8(2):1656–1669, 2024b. doi:10.1109/TETCI.2024.3353674. Chen et al. [2021] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8126–8135, 2021. Yan et al. [2021] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10448–10457, 2021. Mayer et al. [2022] Christoph Mayer, Martin Danelljan, Goutam Bhat, Matthieu Paul, Danda Pani Paudel, Fisher Yu, and Luc Van Gool. Transforming model prediction for tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8731–8740, 2022. Mayer et al. [2024] Christoph Mayer, Martin Danelljan, Ming-Hsuan Yang, Vittorio Ferrari, Luc Van Gool, and Alina Kuznetsova. Beyond sot: Tracking multiple generic objects at once. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6826–6836, 2024. Ye et al. [2022a] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 341–357, Cham, 2022a. Springer Nature Switzerland. ISBN 978-3-031-20047-2. Cui et al. [2024] Yutao Cui, Cheng Jiang, Gangshan Wu, and Limin Wang. Mixformer: End-to-end tracking with iterative mixed attention. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4129–4146, 2024. doi:10.1109/TPAMI.2024.3349519. Wei et al. [2023] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9697–9706, June 2023. Chen et al. [2023] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14572–14581, June 2023. Zheng et al. [2024] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 7588–7596, 2024. Kang et al. [2025] Ben Kang, Xin Chen, Simiao Lai, Yang Liu, Yi Liu, and Dong Wang. Exploring enhanced contextual information for video-level object tracking. In Proceedings of the AAAI conference on artificial intelligence, volume 39, pages 4194–4202, 2025. Cai et al. [2023] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9589–9600, October 2023. Xiao et al. [2022] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 2831–2838, 2022. Hou et al. [2023] Ruichao Hou, Boyue Xu, Tongwei Ren, and Gangshan Wu. Mtnet: Learning modality-aware representation with transformer for rgbt tracking. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1163–1168. IEEE, 2023. Yang et al. [2022] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In Proceedings of the 30th ACM International Conference on Multimedia, M ’22, page 3492–3500, New York, NY, USA, 2022. Association for Computing Machinery. ISBN 9781450392037. doi:10.1145/3503161.3547851. URL https://doi.org/10.1145/3503161.3547851. Zhu et al. [2023] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9516–9526, 2023. Hou et al. [2024] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26551–26561, 2024. Ye et al. [2022b] Junjie Ye, Changhong Fu, Guangze Zheng, Danda Pani Paudel, and Guang Chen. Unsupervised domain adaptation for nighttime aerial tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8896–8905, 2022b. Wei et al. [2024] Haoran Wei, Yanyun Fu, Deyong Wang, Rui Guo, Xueyi Zhao, and Jian Fang. Unsupervised nighttime object tracking based on transformer and domain adaptation fusion network. IEEE Access, 12:130896–130913, 2024. Fu et al. [2024a] Changhong Fu, Yiheng Wang, Liangliang Yao, Guangze Zheng, Haobo Zuo, and Jia Pan. Prompt-driven temporal domain adaptation for nighttime uav tracking. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9706–9713. IEEE, 2024a. Fu et al. [2024b] Changhong Fu, Liangliang Yao, Haobo Zuo, Guangze Zheng, and Jia Pan. Sam-da: Uav tracks anything at night with sam-powered domain adaptation. In 2024 International Conference on Advanced Robotics and Mechatronics (ICARM), pages 31–38, 2024b. doi:10.1109/ICARM62033.2024.10715901. Li et al. [2024] Qiao Li, Kanlun Tan, Qiao Liu, Di Yuan, Xin Li, and Yunpeng Liu. Progressive domain adaptation for thermal infrared object tracking. arXiv preprint arXiv:2407.19430, 2024. Li et al. [2025] Qiao Li, Kanlun Tan, Qiao Liu, Di Yuan, Xin Li, and Yunpeng Liu. Efficient hierarchical domain adaptive thermal infrared tracking. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025. Wu et al. [2024] Yinan Wu, Licheng Jiao, Xu Liu, Fang Liu, Shuyuan Yang, and Lingling Li. Domain adaptation-aware transformer for hyperspectral object tracking. IEEE Transactions on Circuits and Systems for Video Technology, 34(9):8041–8052, 2024. Ganin and Lempitsky [2015] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180–1189. PMLR, 2015. Li et al. [2021] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. IEEE Transactions on Image Processing, 31:392–404, 2021. Loshchilov and Hutter [2017] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. Li et al. [2019] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. Pattern Recognition, 96:106977, 2019. Li et al. [2016] Chenglong Li, Hui Cheng, Shiyi Hu, Xiaobai Liu, Jin Tang, and Liang Lin. Learning collaborative sparse representation for grayscale-thermal tracking. IEEE Transactions on Image Processing, 25(12):5743–5756, 2016. Mayer et al. [2021] Christoph Mayer, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Learning target candidate association to keep track of what not to track. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13444–13454, October 2021. Jiang et al. [2021] Nan Jiang, Kuiran Wang, Xiaoke Peng, Xuehui Yu, Qiang Wang, Junliang Xing, Guorong Li, Guodong Guo, Qixiang Ye, Jianbin Jiao, et al. Anti-uav: A large-scale benchmark for vision-based uav tracking. IEEE Transactions on Multimedia, 25:486–500, 2021.