acceptodds
Under review as a conference paper at ICLR 2027

GeoVLT: Camera Geometry Conditioning for UAV Vision-Language Tracking

Abstract

In UAV-based urban monitoring, continuous target tracking often faces changes in altitude, gimbal pitch angle and focal length, resulting in different camera geometry settings that affect target imaging and scene coverage. Most existing UAV vision-language trackers rely on learned features to implicitly handle the resulting imaging changes, potentially reducing tracking reliability across geometry settings. Camera geometry provides physical cues about these imaging changes, enabling more reliable target modeling. Therefore, we introduce GeoVLT, a vision-language tracking framework that uses these cues throughout the tracking process. Specifically, Geometry-Conditioned Representation first adapts the target representation within a single frame. Geometry-Adaptive Temporal Propagation then maintains historical target information through multi-frame propagation. Geometry-Conditioned Decoding finally combines the single-frame representation with multi-frame historical information for coordinate prediction. To evaluate the use of camera geometry in tracking, we construct GeoUAV, a dataset that provides vision, language, and camera geometry information. For public UAV benchmarks without camera geometry, we train a lightweight estimator on GeoUAV to predict camera geometry from first-frame features. GeoVLT achieves state-of-the-art performance among the evaluated trackers, reaching 69.69% AUC on GeoUAV100 compared with 67.31% for ATCTrack, and 64.11% AUC on UAV20L-NLP compared with 61.66% for UITrack. These results show the effectiveness of explicitly using camera geometry in vision-language tracking. Our code and checkpoints are released at https://anonymous.4open.science/r/GeoVLT.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
UAV Tracking, Vision--Language Tracking, Geometric Information, Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cXM8OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GeoVLT, a UAV vision-language tracking framework that explicitly incorporates camera geometry, including altitude, gimbal pitch angle, and focal length, into target representation, temporal propagation, and coordinate decoding. The paper also introduces GeoUAV, a UAV tracking dataset with language and camera-geometry annotations, and a geometry estimator for benchmarks without such metadata. GeoVLT achieves strong performance on GeoUAV and several public UAV tracking benchmarks.

Critical strengths
  1. The paper explores an interesting and practically relevant direction by explicitly introducing camera geometry into UAV vision-language tracking.

  2. GeoUAV provides camera-geometry annotations that are largely absent from existing UAV tracking datasets, which is potentially useful for future research.

  3. The proposed components consistently improve the baseline in the progressive ablation, and the gains are particularly noticeable under high-altitude settings.

Critical weaknesses
  1. The paper motivates GeoVLT using changes in altitude, gimbal angle, and focal length during UAV tracking, but all camera parameters remain fixed within each sequence in GeoUAV. The estimator is also applied only once using the initial frames and its prediction is reused throughout the sequence. Therefore, the current evaluation does not demonstrate robustness to dynamically changing camera geometry, which is arguably the more realistic and challenging scenario.

  2. The estimator is critical for applying GeoVLT to public datasets without camera metadata, yet the paper does not report direct estimation accuracy for altitude, pitch, or focal length. Consequently, it is difficult to understand how accurate the estimated geometry is and how tracking performance degrades as geometry estimation error increases.

  3. The GeoUAV train/test split appears to contain sequences drawn from the same predefined altitude, pitch, and focal-length ranges. Experiments such as leave-one-altitude/pitch/focal-length-out evaluation or interpolation/extrapolation to unseen configurations would better establish whether GeoVLT learns meaningful geometric relationships rather than dataset-specific correlations.

  4. The authors should mention more related work in somewhere appropriate: ViewSAM: Learning View-aware Cross-modal Semantics for Weakly Supervised Cross-view Referring Multi-Object Tracking, Gen4track: A tuning-free data augmentation framework via self-correcting diffusion model for vision-language tracking, Consistencies are all you need for semi-supervised vision-language tracking, Beyond visual cues: Synchronously exploring target-centric semantics for vision-language tracking.

  5. GCR and GATP consume the full geometry tuple, whereas GCD uses an altitude-conditioned mapping with fixed parameter pairs over altitude intervals. The paper should better justify why pitch and focal length are excluded from the decoding adaptation, particularly given the claim that all three geometry variables affect imaging conditions.

Questions for authors

Please refer to Weaknesses.

5 comments

Dear Reviewer cXM8:

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We provide our responses below.

Reply to W1:

The scope of this study does not include dynamically changing camera geometry within a tracking sequence. Our experimental setting focuses on geometry conditioning across different fixed settings, where camera parameters remain constant within each sequence but vary across sequences. We will clarify this limitation in the revised paper and investigate dynamically changing camera geometry in future work.

Reply to W2:

We have evaluated geometry estimation accuracy on test sequences and compared tracking performance using recorded and estimated geometry.

Geometry estimation MAE:

DatasetAltitude (m)Pitch (°)Focal (×)
GeoUAV308.3112.500.254
GeoUAV6020.4416.020.320
GeoUAV1008.8216.060.304
Overall12.5314.860.293

Tracking AUC (%):

GeometryGeoUAV30GeoUAV60GeoUAV100
Recorded82.5574.5869.69
Estimated82.4872.9567.95

Using estimated geometry reduces AUC by 0.07, 1.63, and 1.74 percentage points compared with recorded geometry. These results quantify the estimator's accuracy and its impact on tracking performance.

Reply to W3:

We retrain GeoVLT after excluding all training sequences with a focal multiplier of 1.5, while retaining the standard model trained on the original training split for comparison. Both models are evaluated on the same 11 test sequences recorded at 1.5. For the holdout model, we further compare the correct focal multiplier (1.5) with an alternative input (1.4), keeping altitude and gimbal pitch angle at their recorded values.

Training ProtocolInput FocalAUCPPn
Standard training1.578.5699.4590.25
Holdout training1.577.7399.2488.55
Holdout training1.474.1794.4684.96

The holdout model achieves 77.73% AUC, only 0.83 points below standard training. Using the correct focal multiplier further improves AUC by 3.56 points over using 1.4, supporting effective geometry conditioning at an unseen focal setting.

Reply to W4:

We have incorporated all four suggested references, including ViewSAM, Gen4Track, Consistencies Are All You Need, and Beyond Visual Cues, into the Introduction and Related Work sections at appropriate locations. These additions provide a more comprehensive overview of recent vision–language tracking research and better clarify the positioning of GeoVLT relative to existing approaches. This will be reflected in the revised version of the paper.

Reply to W5:

GCR and GATP incorporate all three geometry parameters to condition target representation and temporal propagation. GCD, in contrast, performs local aggregation on response and offset maps that have already been conditioned on the full geometry tuple. Since altitude provides a coarse cue for target scale, we use it to select the local decoding parameters, keeping the final readout simple without introducing additional mappings for pitch angle and focal length.

We further compare shared GCD parameter pairs across altitude groups:

GCD ParametersGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Shared (0.10, 0.50)81.9074.1969.69
Shared (0.10, 0.70)82.0374.5867.42
Shared (0.15, 0.70)82.5574.4369.03
Altitude-conditioned GCD82.5574.5869.69

Different altitude groups favor different decoding settings. Altitude-conditioned GCD achieves a mean AUC of 75.61%, compared with 75.34% for the best shared setting among those tested. These results support altitude-dependent decoding adaptation.

Dear Reviewer,

Thank you for your review. The authors have submitted their rebuttal. Could you please take a moment to review their responses?

Thank you for your time and contribution!

Best regards, Area Chair

Most of my concerns have been well addressed.

Reviewer_jYunOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoVLT, a vision-language tracking framework that explicitly incorporates camera geometry into UAV tracking. The framework considers three geometry factors, including altitude, gimbal pitch angle, and focal length, and introduces Geometry-Conditioned Representation (GCR), Geometry-Adaptive Temporal Propagation (GATP), and Geometry-Conditioned Decoding (GCD) to incorporate geometry information into visual representation, temporal propagation, and coordinate prediction. The paper also introduces GeoUAV, a UAV vision-language tracking dataset containing 495 sequences, 56K frames, eight target classes, and controlled camera geometry settings. In addition, a geometry estimator predicts camera geometry for public datasets that do not provide geometry annotations. Experiments show that GeoVLT achieves consistent improvements over existing vision-only and vision-language trackers on GeoUAV and several public benchmarks.

Critical strengths
  1. The paper addresses an important and relatively underexplored factor in UAV vision-language tracking. The paper explicitly identifies camera geometry as an important factor that affects UAV tracking performance. Modeling altitude, gimbal pitch angle, and focal length provides a meaningful way to connect the physical imaging configuration with visual target appearance and localization. The reported results show consistent improvements across different altitude subsets, indicating that geometry information provides useful cues for UAV tracking. The research direction is also relevant to practical UAV tracking scenarios, where different imaging configurations can substantially affect target scale and appearance.

  2. The GeoUAV dataset provides a useful basis for studying geometry-aware tracking. GeoUAV jointly provides tracking sequences, language descriptions, and controlled camera geometry settings. The dataset varies altitude, gimbal pitch angle, and focal length across sequences, which allows the authors to analyze tracking performance under different imaging configurations. The explicit geometry information also enables experiments that are difficult to conduct on existing datasets without camera geometry annotations. Therefore, GeoUAV provides a potentially useful resource for studying the relationship between camera geometry and UAV tracking performance.

  3. The paper provides a relatively comprehensive experimental evaluation. The paper evaluates GeoVLT on both the proposed GeoUAV dataset and several public benchmarks. It also includes progressive ablation studies and analyses of different geometry configurations. The reported results show that the proposed method consistently improves over the compared baselines under multiple experimental settings. The geometry estimation module further enables the framework to operate on datasets without explicit geometry annotations, which increases the potential applicability of the approach.

Critical weaknesses
  1. The main experimental protocol does not fully support the paper's central motivation regarding changing camera geometry. The paper motivates GeoVLT from the perspective that UAV tracking needs to handle changing camera geometry. However, the camera settings remain fixed within each GeoUAV sequence and only vary across different sequences. Therefore, the main benchmark primarily evaluates the ability to adapt to different fixed imaging configurations rather than the ability to handle geometry changes during a tracking trajectory. This distinction is important because GATP is designed to model geometry-aware temporal propagation, while the current benchmark does not directly test whether the method remains effective when the geometry changes over time. As a result, the experimental setting does not fully match the central motivation of the proposed framework.

  2. The contribution of explicit geometry is not sufficiently isolated from the additional architectural capacity. The proposed framework introduces several geometry-conditioned components, including GCR, GATP, and GCD. Although the progressive ablation shows that adding these components improves the overall performance, the current experiments do not completely separate the benefit of geometry information from the benefit of introducing additional network structures and conditioning mechanisms. For example, it remains unclear whether a similarly designed architecture with another auxiliary signal or a comparable increase in model capacity achieves similar improvements. More controlled comparisons are therefore needed to establish that the performance gains specifically result from meaningful camera geometry rather than increased model capacity or architectural specialization.

  3. The geometry estimation component introduces an additional source of uncertainty that is not sufficiently analyzed. The method relies on estimated camera geometry when it is applied to public datasets without geometry annotations. However, the tracking performance can depend on the accuracy of the geometry estimator. The paper does not provide a sufficiently detailed analysis of the relationship between geometry estimation error and tracking performance. In particular, it remains unclear how sensitive GeoVLT is to inaccurate altitude, pitch angle, or focal length estimates, and whether the reported gains remain stable when the estimated geometry deviates from the true geometry. This issue is important for evaluating the robustness of the proposed framework in practical scenarios.

  4. The experimental design does not fully demonstrate the necessity of modeling all three geometry factors. The framework jointly considers altitude, gimbal pitch angle, and focal length, but the experiments provide limited evidence about the individual importance of these factors. Since these parameters can affect the visual input in different ways, it is useful to understand whether each factor contributes independently and whether some factors dominate the observed improvements. A more systematic factor-wise analysis can make the motivation for the complete geometry representation more convincing.

Questions for authors
  1. Will the complete dataset and source code be publicly released?

  2. Can the authors evaluate GeoVLT on sequences where camera geometry changes within a tracking sequence?

  3. Can the authors better isolate the contribution of geometry information from additional model capacity?

  4. How sensitive is GeoVLT to geometry estimation errors?

  5. What are the individual contributions of altitude, pitch angle, and focal length?

  6. What additional benefit does GATP provide when geometry remains constant within a sequence?

6 comments

Dear Reviewer jYun:

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We provide our responses below.

Reply to W1 & Q2:

The scope of this study does not include dynamically changing camera geometry within a tracking sequence. Our experimental setting focuses on geometry conditioning across different fixed imaging configurations, where camera parameters remain constant within each sequence but vary across sequences.

Although camera geometry remains fixed within a sequence, GATP uses it to condition temporal memory propagation, while target observations, tracking confidence, and memory evolve over time. We will clarify this limitation in the revised manuscript and investigate dynamically changing camera geometry in future work.

Reply to W2 & Q3:

To better distinguish the contribution of camera geometry from additional model capacity, we provide two complementary comparisons.

1. Capacity-matched Comparison. We construct a parameter-matched baseline by replacing the proposed geometry-conditioning mechanisms with a simple residual MLP receiving the same geometry inputs. The two models differ by only 10,817 parameters, corresponding to 0.0042% of the total parameters and 0.0080% of the trainable parameters.

Parameter GroupSimple GeometryGeoVLTDifference
Geometry Fusion43,879,10443,889,92110,817 (0.0246%)
Full Model258,710,853258,721,67010,817 (0.0042%)
Trainable Parameters135,060,549135,071,36610,817 (0.0080%)
MethodGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Simple Geometry81.7372.8166.84
GeoVLT82.5574.5869.69

GeoVLT consistently improves AUC by 0.82, 1.77, and 2.85 percentage points, respectively. These results indicate that the proposed conditioning interfaces offer benefits beyond simple geometry fusion, despite nearly identical model capacity.

2. Geometry Input Validity. We further examine the existing geometry-input controls, which maintain the same model architecture and parameters while varying the geometry inputs.

Geometry InputGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Correct82.5574.5869.69
Fixed82.3573.4366.14
Shuffled81.0173.0867.80

Correct geometry consistently outperforms fixed and shuffled inputs, providing further evidence that the gains depend on meaningful camera geometry rather than model capacity alone.

Reply to W3 & Q4:

We evaluate geometry estimation accuracy on test sequences and analyze its impact on tracking performance.

1. Geometry Estimation Accuracy. The mean absolute errors (MAEs) are reported below.

DatasetAltitude (m)Pitch (°)Focal (×)
GeoUAV308.3112.500.254
GeoUAV6020.4416.020.320
GeoUAV1008.8216.060.304
Overall12.5314.860.293

2. Impact on Tracking Performance. We compare tracking using recorded and estimated geometry on the same sequences.

Geometry InputGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Recorded82.5574.5869.69
Estimated82.4872.9567.95

Using estimated geometry reduces AUC by 0.07, 1.63, and 1.74 percentage points relative to recorded geometry, indicating that much of the tracking performance is retained despite estimation errors.

Reply to W4 & Q5:

To investigate the role of each geometry factor, we conduct an inference-time sensitivity analysis by fixing one parameter at a time while keeping the other two at their recorded values.

Geometry InputGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Full Geometry82.5574.5869.69
Fixed Altitude (60 m)82.3574.5868.30
Fixed Pitch (60°)82.1474.4867.27
Fixed Focal (1.5×)82.0973.3466.24

Replacing individual geometry factors generally reduces tracking performance, with the largest degradation observed when fixing focal length. On GeoUAV100, fixing altitude, gimbal pitch angle, and focal length decreases AUC by 1.39, 2.42, and 3.45 percentage points, respectively.

These results indicate different sensitivities to the three geometry inputs. Fixing altitude at 60 m leaves GeoUAV60 unchanged because its recorded altitude is also 60 m.

Reply to Q1:

The complete GeoUAV dataset and source code are not yet publicly released. We will release both after acceptance of the paper to facilitate reproducibility and future research.

Reply to Q6:

Although camera geometry remains fixed within each sequence, target observations, tracking confidence, and temporal memory evolve across frames. GATP uses sequence-level geometry to condition how historical target information is propagated.

Gating StrategyGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Confidence-only82.1174.1267.86
GeoVLT82.5574.5869.69

To further validate this design, we compare geometry-conditioned memory gating with a confidence-only alternative. The geometry-conditioned variant improves AUC by 0.44, 0.46, and 1.83 percentage points on GeoUAV30, GeoUAV60, and GeoUAV100, respectively.

These results demonstrate that sequence-level geometry provides useful context for temporal memory propagation, even when camera parameters remain unchanged within a sequence.

Dear Reviewer,

Thank you for your review. The authors have submitted their rebuttal. Could you please take a moment to review their responses?

Thank you for your time and contribution!

Best regards, Area Chair

Reviewer_RJtiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents GeoVLT, a UAV vision-language tracker that uses altitude, gimbal pitch angle, and focal length to condition representations, temporal memory, and coordinate decoding through GCR, GATP, and GCD. Camera parameters are fixed within each sequence. The authors introduce GeoUAV, containing 495 sequences and approximately 56K frames with tracking, language, and camera-geometry annotations. For public benchmarks without measured geometry, a learned estimator predicts the parameters once from initial visual features. The paper reports AUC improvements across three altitude subsets and four public benchmark settings over the methods included in the comparisons.

Critical strengths
  1. The paper studies a relevant and interpretable source of variation in aerial tracking: the acquisition geometry affects target scale, visible detail, and perspective. GeoUAV pairs tracking and language annotations with recorded camera settings under controlled conditions, providing a useful resource for studying this factor. The combination of a focused research question and a benchmark with the corresponding metadata is a meaningful contribution.

  2. GeoVLT incorporates geometry coherently into representation, temporal propagation, and localization. Table 3 reports improvements of 2.84, 3.41, and 4.07 AUC points over the ablation baseline on GeoUAV30, GeoUAV60, and GeoUAV100, respectively. Table 4 further supports the value of matched geometry, particularly on GeoUAV100, where recorded geometry achieves 69.69 AUC compared with 66.14 for a fixed tuple. Public-benchmark experiments demonstrate that the overall approach can also be applied with estimated geometry. Taken together, the dataset and the empirical evidence support a useful contribution and motivate my weak-acceptance recommendation.

Critical weaknesses
  1. The method builds on established prompting, feature-modulation, and memory mechanisms, and the analysis could distinguish the contribution of geometry conditioning from additional capacity more precisely. The progressive ablations and geometry-input controls are informative, but comparisons against simpler alternatives receiving the same geometry inputs would strengthen the justification for the chosen interfaces. Confidence-only memory gating and a shared GCD parameter pair across altitude groups would also help isolate the benefits of geometry-dependent adaptation. This limits the precision of the mechanism-level conclusions, while the reported overall gains remain encouraging.

  2. Several evaluation details need clarification. Appendix B tests 22 decoder parameter pairs per altitude group, but an independent validation protocol for parameter selection is not described. The training or fine-tuning settings of the baselines and the omission of AVLTrack and ATCTrack from Table 5 should also be explained. In addition, the DTB70-NLP gain over the best listed comparator is only 0.39 AUC points (54.80 versus 54.41), so uncertainty estimates would help assess this particular margin. I regard these as important clarification and validation issues that should be addressed to make the empirical claims more precise.

Questions for authors
  1. Protocol clarification: Which split was used to select the temporal unroll length and the decoder parameters? Please specify the altitude interval boundaries, how settings are assigned to estimated heights on public benchmarks, and whether test sequences were excluded from all model and hyperparameter selection.
  2. Evidence request: Under matched training data and comparable model capacity, how does the proposed conditioning compare with a simpler geometry-input baseline? Results for confidence-only memory gating or a shared GCD setting would help explain the contribution of the geometry-dependent interfaces.
  3. Comparison clarification: Were the baselines trained or fine-tuned on the same GeoUAV training split, and which pretrained checkpoints were used? Please explain the absence of AVLTrack and ATCTrack from Table 5 and, where available, provide comparable results and an uncertainty assessment for small margins.
  4. Estimator clarification: What are the geometry-estimation errors on held-out GeoUAV sequences, and how does tracking with estimated geometry compare with recorded geometry and a fixed tuple on the same sequences? An ablation of token-level distillation would further clarify the estimator’s contribution.
6 comments

Dear Reviewer RJti:

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We provide our responses below.

Reply to W1 & Q2:

To better distinguish the benefits of geometry conditioning from additional model capacity and simpler alternatives, we conduct three controlled comparisons.

1. Simple Geometry-input Baseline. We construct a parameter-matched baseline by replacing the proposed geometry-conditioning mechanisms with a simple residual MLP receiving the same geometry inputs. The two models differ by only 10,817 parameters, corresponding to 0.0042% of the total parameters and 0.0080% of the trainable parameters.

Parameter GroupSimple GeometryGeoVLTDifference
Geometry Fusion43,879,10443,889,92110,817 (0.0246%)
Full Model258,710,853258,721,67010,817 (0.0042%)
Trainable Parameters135,060,549135,071,36610,817 (0.0080%)
MethodGeoUAV30 AUC / P / PnGeoUAV60 AUC / P / PnGeoUAV100 AUC / P / Pn
Simple Geometry81.73 / 96.55 / 93.9372.81 / 89.98 / 84.9166.84 / 89.70 / 77.30
GeoVLT82.55 / 96.81 / 94.8474.58 / 92.27 / 87.8169.69 / 93.15 / 78.71

With nearly identical model capacity and the same geometry inputs, GeoVLT improves AUC by 0.82, 1.77, and 2.85 percentage points, respectively. These results support the effectiveness of the proposed conditioning interfaces beyond simple geometry fusion or increased parameter capacity.

2. Confidence-only Memory Gating. We compare geometry-conditioned memory gating with a confidence-only alternative that removes geometry information from the gating function.

Gating StrategyGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Confidence-only82.1174.1267.86
GeoVLT82.5574.5869.69

Geometry-conditioned gating improves AUC by 0.44, 0.46, and 1.83 percentage points, respectively. These results indicate that camera geometry provides additional benefits for temporal memory propagation beyond tracking confidence alone.

3. Shared GCD Parameters. We evaluate three shared GCD parameter pairs, each applied uniformly across all altitude groups.

GCD Parameters (τ, ρ)GeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Shared (0.10, 0.50)81.9074.1969.69
Shared (0.10, 0.70)82.0374.5867.42
Shared (0.15, 0.70)82.5574.4369.03
Altitude-conditioned GCD82.5574.5869.69

Different altitude groups favor different decoding settings. Together, these comparisons provide additional evidence for the proposed geometry-conditioning design.

Reply to W2 & Q1 & Q3:

We provide the following clarifications regarding the evaluation protocol, baseline comparisons, and cross-dataset results.

1. Hyperparameter Selection. GeoUAV is divided into 396 training sequences and 99 test sequences. We train GeoVLT on all 396 training sequences and evaluate different parameter settings on the training set to select the well-behaved hyperparameter. The test sequences are not involved in model training or hyperparameter selection.

2. Altitude Interval Mapping. GCD uses the following predefined altitude intervals:

Altitude (h)GCD Parameters (τ, ρ)
h < 45 m(0.15, 0.70)
45 ≤ h < 80 m(0.10, 0.70)
h ≥ 80 m(0.10, 0.50)

For public benchmarks without geometry annotations, the estimator predicts altitude from the initial frames, and the same interval mapping is applied. The selected parameters remain fixed within each sequence.

3. Baseline Training Protocol. We initialize each baseline using its officially released pretrained backbone rather than the authors' full tracking checkpoint. All baselines are trained on the same GeoUAV split for 40 epochs.

4. AVLTrack and ATCTrack. Due to limited time and computational resources, we have not yet completed the training and evaluation of AVLTrack and ATCTrack. However, we are currently conducting these experiments and will incorporate the results into the paper once training and evaluation are completed.

5. DTB70-NLP Performance. We conducted multiple evaluations on DTB70-NLP and consistently observed that GeoVLT outperforms the baseline, with AUC improvements of less than 1% across repeated tests. Although the performance gain is relatively modest, the consistent improvements demonstrate the stability of GeoVLT on this benchmark and further support its cross-dataset generalization capability.

Reply to Q4:

We provide additional experiments to evaluate geometry estimation accuracy, its impact on tracking performance, and the contribution of token-level distillation.

1. Geometry Estimation Accuracy. We evaluate estimation errors on test sequences.

DatasetAltitude (m)Pitch (°)Focal (×)
GeoUAV308.3112.500.254
GeoUAV6020.4416.020.320
GeoUAV1008.8216.060.304
Overall12.5314.860.293

2. Tracking with Estimated Geometry. We compare recorded and estimated on the same test sequences.

Geometry InputGeoUAV30 AUCGeoUAV60 AUCGeoUAV100 AUC
Recorded82.5574.5869.69
Estimated82.4872.9567.95
Estimated w/o Distillation82.1972.7767.44

3. Token-level Distillation Ablation. Using the same tracker, we evaluate an estimator trained without token-level distillation, as included in the comparison above.

Using estimated geometry reduces AUC by 0.07, 1.63, and 1.74 percentage points compared with recorded geometry, indicating that much of the tracking performance is retained despite estimation errors.

Removing token-level distillation further reduces AUC by 0.29, 0.18, and 0.51 percentage points, respectively. These AUC improvements support the contribution of token-level distillation to downstream tracking performance.

Dear Reviewer,

Thank you for your review. The authors have submitted their rebuttal. Could you please take a moment to review their responses?

Thank you for your time and contribution!

Best regards, Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.