acceptodds
Under review as a conference paper at ICLR 2027

Temporal Geometry-Aware Local Search for Thermal Infrared Anti-UAV Tracking

Abstract

Thermal infrared (TIR) UAV tracking is challenged by small targets, weak appearance cues, and large inter-frame displacement. For bounded local trackers that restrict the search to a crop around the previous target estimate, once the target leaves the search crop, appearance modeling alone cannot recover the missing visual evidence. Expanding the search region may introduce more background clutter, while ambiguous local cues make it difficult to decide where to search. We therefore view bounded recovery as two distinct decisions: where to acquire additional evidence and which hypothesis to trust. Based on this view, we propose TGTrack, which decouples search allocation from visual selection. A time-aware Temporal Geometry Branch models sparsely sampled trajectory states in global coordinates by encoding center and scale evolution, visual confidence, and temporal information, and predicts target geometry and localization reliability. When the predicted center leaves the normal search support, TGTrack allocates one bounded geometry-guided Recrop. We further introduce Gate–Hann, which modulates the within-Recrop spatial prior according to predicted reliability. Final cross-region selection remains appearance-driven using raw visual confidence. Extensive experiments show that TGTrack achieves state-of-the-art results among the compared local-search trackers, improves the controlled LoRAT-L baseline by 2.86 AUC points on AntiUAV410 while running at 292.01 FPS, and by 4.03 points on AntiUAV under direct transfer without target-domain adaptation. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
visual object tracking, thermal infrared tracking, anti-UAV tracking, temporal modeling, trajectory modeling, search allocation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_ocv1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TGTrack augments LoRAT-L with a temporal geometry branch that predicts target location, scale, and reliability from sparse trajectory histories. Geometry conditionally allocates an additional search crop, Gate–Hann adjusts its spatial prior, and raw visual confidence selects between regional hypotheses. Evaluation covers AntiUAV410 and direct transfer to AntiUAV.

Critical strengths
  1. Geometry determines where to acquire additional evidence, while appearance determines the final hypothesis. Reliability-dependent Hann modulation is consistent with uncertainty about the predicted crop center.

  2. TGTrack improves the matched LoRAT-L baseline by 2.86 AUC points, with a positive sequence-level bootstrap confidence interval. The 4.03-point improvement under direct cross-dataset transfer further supports practical value.

  3. Constant-velocity and generic trajectory controls, memory sensitivity, gate diagnostics, and cross-region selection analysis help explain the gains. The reported computation accounts for the observed conditional Recrop frequency.

Critical weaknesses
  1. The motion-model controls are useful, but different policies may trigger additional searches at different frequencies. Comparisons against simpler search policies under matched execution rates or computation are needed.

  2. The proposed branch combines age encoding, confidence weighting, query pooling, horizon fusion, and motion correction. The generic Transformer comparison changes several elements together, leaving their individual necessity unclear.

  3. The reported reliability analysis uses Stage-1 cached histories, whereas Gate–Hann matters specifically on executed Recrop frames. Its behavior on these harder, policy-dependent cases needs separate validation.

  4. Evaluation uses two related benchmarks and one visual backbone. The tracker always outputs a box, while legacy precision rewards absent frames; although acknowledged, this limits conclusions about prolonged disappearance and recovery reliability.

Questions for authors
  1. Do the gains persist against simpler additional-search policies under matched Recrop frequency or computational cost?

  2. Which temporal-branch components contribute most, beyond the generic Transformer baseline?

  3. How reliable is the gate specifically on executed Recrop frames and after long gaps without admitted observations?

  4. Can you report visibility-conditioned results and validate transfer to another benchmark or visual backbone?

2 comments

Q1 (W1). Comparison under comparable Recrop budgets

We thank the reviewer for recognizing the value of our decoupled search-allocation design. We address the four questions with additional budget-controlled comparisons, component ablations, online gate analysis, and visibility-conditioned and transfer evaluations.

To separate search-allocation quality from additional search frequency, we reevaluate the geometry strategies on AntiUAV410 under a common protocol. The learnable geometry models are independently retrained. All strategies use the same visual tracker, conditional Recrop mechanism, and fixed Hann weight (λh=0.45\lambda_h=0.45), without Gate–Hann. For the budget-controlled variants, eligible Recrop requests are periodically subsampled with a fixed per-sequence phase.

Search allocationAUCSARecrop (%)Executions
Constant Velocity63.3964.511.83402,376
Generic Learnable Geometry63.6364.762.09202,711
Proposed (Low Budget)64.8966.030.97091,258
Proposed (Near-Matched)64.9766.111.87622,431
Proposed (Unrestricted)65.3566.499.609012,450

At 1.8762% Recrop, the proposed strategy gains 1.58 AUC points over Constant Velocity (1.8340%), with only 55 additional executions. It also gains 1.34 points over Generic Learnable Geometry while executing 280 fewer Recrops.

Even at a 0.9709% Recrop rate, it achieves 64.89 AUC, exceeding both alternatives. The results support an advantage in search allocation that is not solely attributable to more frequent additional visual searches.


Q2 (W2). Contributions of individual temporal components

The generic Transformer comparison changes several design choices together. We therefore evaluate six variants of full TGTrack, each removing one component, to examine their contributions separately.

ConfigurationAUC (%)Δ\DeltaAUC
Full TGTrack66.31--
w/o Age Embedding65.17-1.14
w/o Relative-Age Bias65.79-0.52
w/o Confidence Weighting65.33-0.98
w/o Query Pooling65.37-0.94
w/o Learned Horizon Fusion64.36-1.95
w/o Adaptive Motion Correction65.52-0.79

Every removal lowers AUC in these runs. Learned Horizon Fusion has the largest effect (-1.95 points), followed by Age Embedding (-1.14) and Confidence Weighting (-0.98). Relative-Age Bias, Query Pooling, and Adaptive Motion Correction also contribute, with drops of 0.52, 0.94, and 0.79 points.

These results support explicit modeling of elapsed time, observation reliability, and historical motion beyond the generic trajectory Transformer.

Q3 (W3). Gate reliability during online Recrop and long observation gaps

We thank the reviewer for distinguishing cached-history analysis from online Recrop behavior. We analyze predicted gate values and center errors on valid online Recrop predictions, using the following gate intervals.

Gate intervalFramesMean gateMean center error (px)
[0.2,0.4)[0.2,0.4)7340.3299253.82
[0.4,0.6)[0.4,0.6)1,0050.5071214.57
[0.6,0.8)[0.6,0.8)1,4420.7094187.81
[0.8,1.0][0.8,1.0]3,8680.886898.64
Overall7,0490.7384149.57

Higher gate intervals have lower mean center error, decreasing from 253.82 to 98.64 px. This provides online, group-level evidence that the gate carries information about localization accuracy.

Long observation gaps. We also compare mean gates with the GT-based reliability target used in training, grouping predictions by the elapsed time since the latest admitted trajectory state.

Observation gapMean gateMean GT target
2–5 frames0.6580.537
6–15 frames0.6530.553
$\geq$16 frames0.5650.616

Mean gate decreases from 0.658 to 0.565 as the gap increases, indicating more conservative spatial regularization after prolonged periods without admitted observations. As further quantified in the appendix, the predicted gate exhibits a strong Spearman correlation with localization accuracy (ρ(gt,−et)=0.7004\rho(g_t,-e_t)=0.7004), supporting its role as a reliability indicator.

The learned gate serves as a relative reliability indicator for modulating the Recrop spatial prior, rather than an exact estimate of geometric prediction accuracy.


Q4 (W4). Visibility-conditioned evaluation and generalization

To assess visibility-conditioned performance and broader applicability, we provide additional evaluations on visible-target frames, two alternative tracking architectures, and zero-shot transfer to NfS-30.

Visibility-conditioned evaluation. We compare LoRAT-L and TGTrack on the full test set and on frames with valid visible targets.

SubsetBase AUCTG AUCBase SATG SA
All frames63.4566.3164.5667.47
Visible (valid GT)64.3967.4165.5268.58

Of 129,691 frames, 2,622 are target-absent (exist=0) and 41 have invalid GT boxes. Under this evaluation, they remain in the all-frame calculation with zero AUC and SA contributions. Metrics are averaged per sequence. On the remaining 127,028 valid-GT frames, TGTrack gains 3.02 AUC and 3.06 SA points. The gains therefore persist for visible-target localization.

Cross-framework evaluation. We evaluate integrations with OSTrack-256 and LETrack on AntiUAV410, comparing each augmented tracker with its own baseline.

Tracking frameworkBaseline AUCAugmented AUCΔ\DeltaAUC
LoRAT-L63.4566.31+2.86
OSTrack-25656.0757.21+1.14
LETrack57.1159.65+2.54

The gains of 1.14 AUC on OSTrack-256 and 2.54 AUC on LETrack support applicability beyond LoRAT-L.

Zero-shot cross-dataset transfer. On NfS-30, we directly reuse the AntiUAV410-trained model and inference parameters, without retraining or tuning.

MethodAUCPrecisionNormPOP50OP75
LoRAT-L60.4473.7675.1275.1050.33
TGTrack63.2276.1977.9878.5252.73
Δ\Delta+2.78+2.43+2.86+3.42+2.40

TGTrack gains 2.78 AUC points and improves all reported NfS-30 metrics. Together with the original +4.03 AUC transfer gain on AntiUAV, these results support applicability across the evaluated architectures and datasets.

We will include these results and clarifications in the revised manuscript. Many thanks for your constructive feedback.

Best regards,
On behalf of the authors

Reviewer_AekAOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TGTrack, a temporal geometry-aware local search framework for thermal infrared anti-UAV tracking. The main idea is to decouple search allocation from appearance-based hypothesis selection. A temporal geometry branch predicts the target center, scale, and localization reliability from historical trajectory states, while a conditional geometry-guided Recrop provides additional visual evidence when the predicted center falls outside the normal search region. The method achieves promising performance on AntiUAV410 and demonstrates favorable efficiency compared with persistent search-region enlargement.

Critical strengths

[1] The paper addresses an important limitation of local-search trackers: appearance modeling alone cannot recover targets that have moved outside the search region. [2] Decoupling geometry-guided search allocation from appearance-based selection is a reasonable and practically useful formulation. [3] The method incorporates observation confidence and irregular temporal intervals into trajectory modeling. [4] The experimental results demonstrate consistent improvements over the controlled LoRAT-L baseline, with relatively low additional computational cost.

Critical weaknesses
  1. Insufficient evaluation of temporal geometry prediction, In particular, there is a lack of assessment regarding the correctness of the geometric branches.
  2. The source of the improvement from Gate–Hann is unclear, whether the fixed-weight approach is reasonable requires further verification.
Questions for authors

[1] Although the proposed geometry branch improves the final tracking AUC, the authors provide limited systematic analysis of its prediction accuracy and recovery effectiveness under different motion conditions. In particular, how does the center prediction error vary with the prediction horizon, abrupt motion changes, and prolonged occlusions? The authors are encouraged to report prediction errors, trigger recall, Recrop coverage, and recovery success rates under different challenging conditions. [2] How to evaluate the effectiveness of geometric branch prediction trajectories, and the impact of branch prediction accuracy on the performance of the tracking process. [3] The proposed Gate–Hann mechanism reduces the effective Hann weight according to predicted localization reliability. However, the observed improvement may partly result from reducing the Hann weight rather than reliability-aware adaptation itself. The authors should compare Gate–Hann with multiple fixed Hann weights, constant or randomized gates, and ideally an oracle reliability gate. These comparisons are important for isolating the actual contribution of reliability modeling.

2 comments

We thank the reviewer for recognizing the motivation and practical value of TGTrack and for suggesting more direct evaluations of geometric prediction and Gate–Hann. We address prediction behavior under challenging conditions, its downstream tracking impact, and the contribution of reliability-adaptive spatial weighting.

Q1 (W1). Geometric prediction under challenging conditions

We analyze center prediction error using closed-loop tracking histories on AntiUAV410. The prediction horizon is the number of frames since the last admitted observation; ground truth is used only for evaluation.

Observation HorizonValid PredictionsMean Center Error (px)Final ΔIoU
2–5 frames5,312164.98+0.01427
6–15 frames3,415186.30+0.01307
≥16 frames3,115716.36+0.00851

Mean center error increases from 164.98 to 716.36 px with longer observation gaps, yet executed Recrops retain positive final IoU gains relative to normal-search candidates. Under Fast Motion (FM), 14,260 valid predictions yield a mean center error of 189.49 px, while out-of-search Recrops improve final IoU by +0.00323 on average. These results suggest that geometry-guided Recrop can provide useful visual evidence even when prediction accuracy degrades.

Recovery after occlusion. On the same annotated occlusion events and identical post-occlusion observation windows, we measure recovery when the final predicted box reaches IoU ≥ 0.5 at least once after reappearance. We also report first-frame localization separately.

OcclusionEventsBase recoveryTG recoveryFirst frame (B / T)
1–9 frames1154.55%72.73%5/11 vs. 5/11
10–29 frames540.00%60.00%2/5 vs. 2/5
≥30 frames616.67%83.33%1/6 vs. 0/6
Overall2240.91%72.73%8/22 vs. 7/22

Here B and T denote LoRAT-L and TGTrack. Across 22 events, recovery improves from 40.91% to 72.73% (+31.82 percentage points). For occlusions of at least 30 frames, TGTrack recovers in 5/6 events within the observation window, although it does not localize the target on the first frame after reappearance.


Q2 (W1). Impact of geometric prediction on tracking

TGTrack uses predicted geometry to acquire additional visual evidence, not to output the final target box directly. We therefore examine how center error relates to Recrop coverage and to the IoU changes of the additional visual candidate and final output.

Center errorRecrop coverage (%)Recrop ΔIoUFinal output ΔIoU
0 ≤ e < 10 px100.00+0.25801+0.18390
10 ≤ e < 30 px88.68+0.28222+0.23646
30 ≤ e < 100 px67.94+0.06047+0.05073
e ≥ 100 px33.09+0.00654+0.00594

When center error is below 30 px, Recrop coverage is 88.68–100.00%, accompanied by substantial gains in the Recrop candidate and final output. For errors ≥100 px, coverage falls to 33.09% and final ΔIoU is only +0.00594. This illustrates how prediction quality affects the opportunity for visual recovery. Even with large errors, the final gain remains slightly positive, consistent with appearance-based selection not automatically accepting the Recrop hypothesis.

Near-budget search comparison. We also compare geometry strategies under the same LoRAT-L tracker, conditional Recrop rule, and fixed Hann (λₕ = 0.45), without Gate–Hann. For the proposed budget variants, eligible Recrop requests are periodically subsampled with a fixed per-sequence phase, without using ground truth or candidate quality.

Search allocationAUCSARecrop (%)Executions
Constant Velocity63.3964.511.83402,376
Generic Learnable Geometry63.6364.762.09202,711
Proposed (Low Budget)64.8966.030.97091,258
Proposed (Near-Matched)64.9766.111.87622,431
Proposed (Full)65.3566.499.609012,450

At nearly the same Recrop rate as Constant Velocity (1.8762% vs. 1.8340%), the proposed strategy gains +1.58 AUC with 55 more executions. It also exceeds Generic Learnable Geometry by +1.34 AUC despite 280 fewer Recrops. Thus, the gains are not solely attributable to a higher search frequency.

Component ablations give complementary evidence: removing Horizon Fusion, Age Embedding, or Confidence Weighting decreases full TGTrack AUC by 1.95, 1.14, and 0.98 points, respectively.

Q3 (W2). Gate–Hann beyond reducing a fixed spatial prior

We agree that comparison with the default Hann weight alone cannot isolate reliability-aware adaptation. We therefore evaluate multiple fixed weights and a shuffled-gate control, which approximately preserves the gate-value distribution while disrupting its assignment to individual frames.

Spatial priorAUC (%)SA (%)
Fixed Hann, λₕ = 0.065.5066.64
Fixed Hann, λₕ = 0.165.7366.87
Fixed Hann, λₕ = 0.265.9467.09
Fixed Hann, λₕ = 0.365.9867.13
Fixed Hann, λₕ = 0.465.6666.81
Shuffled Gate65.9667.12
Gate–Hann (Ours)66.3167.47

Gate–Hann improves on the best tested fixed weight (λₕ = 0.3) by +0.33 AUC and +0.34 SA, and on Shuffled Gate by +0.35 AUC and +0.35 SA. A constant gate gₜ = c yields a fixed effective Hann weight cλₕ; thus, the fixed-weight sweep also serves as a constant-gate control. These comparisons suggest that frame-specific reliability alignment matters beyond simply reducing or varying Hann weight.

Gate–error relationship.

Gate intervalFramesMean center error (px)
[0.2, 0.4)734253.82
[0.4, 0.6)1,005214.57
[0.6, 0.8)1,442187.81
[0.8, 1.0]3,86898.64

Mean center error decreases as the predicted gate increases, consistent with the original cached-history Spearman correlation of ρ(gₜ, −eₜ) = 0.7004. These results further support the learned gate as a meaningful relative indicator of geometric prediction reliability. It adjusts only the within-Recrop spatial prior; the final output remains appearance-driven.

Many thanks for the constructive suggestions.

Best regards,
On behalf of the authors

Reviewer_WB9eOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the problem of the target moving outside the search region in anti-UAV tracking (a subfield of SOT) . It proposes an adaptive search strategy that uses temporal information to predict potential search regions, aiming to perform better than simply enlarging the search region. The method is evaluated on AntiUAV410 and AntiUAV.

Critical strengths
  1. Targets moving outside the search region, whose size is determined by a manually set search-region ratio, is a longstanding and unresolved problem in single object tracking. The paper attempts to address this important issue.

  2. Gate–Hann and the temporal modeling have clear engineering motivations.

Critical weaknesses
  1. The paper does not sufficiently explain why the method is specifically designed for anti-UAV tracking. Targets moving outside the search region is a general challenge in general video tracking or SOT. Small objects are also challenging in general SOT and UAV tracking, where the goal is to track objects in videos captured by UAVs. The authors do not adequately explain why the method would not apply to these settings.

  2. Many existing methods use re-detection to recover targets that move outside the search region such as (Effective Local and Global Search for Fast Long-Term Tracking) and Global track in the paper. Prior work on predicting future target locations through temporal modeling, such as (PVT++: A Simple End-to-End Latency-Aware Visual Tracking Framework), is also not sufficiently discussed. Overall, I feel the insights beyond existing work are limited.

  3. Regarding efficiency, the method reduces FPS from 349 to 292, introducing substantial overhead, especially for applications with strict real-time requirements.

Questions for authors

Please refers to weakness

7 comments

We sincerely thank you for recognizing the practical motivation and consistent improvements of TGTrack, as well as for your constructive feedback.

W1. Applicability beyond thermal anti-UAV tracking

We agree that targets can leave the search region in general single-object tracking (SOT), including in footage from UAV-mounted cameras. We chose thermal infrared anti-UAV tracking as a particularly challenging case because small targets, weak appearance cues, and sudden displacement often occur together, making reacquisition harder. TGTrack guides the additional search using box histories, frame intervals, and tracking confidence, which are also available from other local trackers.

Following your suggestion, we evaluated the AntiUAV410-trained TGTrack directly on NfS-30, a general SOT benchmark, without retraining or hyperparameter tuning.

MethodAUCPrecisionNormalized precisionOP50OP75
LoRAT-L60.4473.7675.1275.1050.33
TGTrack63.2276.1977.9878.5252.73
Change+2.78+2.43+2.86+3.42+2.40

TGTrack improves AUC by 2.78 points and overlap precision at both thresholds. We will clarify the scope of TGTrack and include these results in the revised manuscript.


W2. Relation to global search and temporal prediction

Thank you for recognizing the motivation and consistent gains of TGTrack. TGTrack separates search allocation from target selection. When the predicted center falls outside the normal crop, TGTrack uses the predicted geometry to place one additional bounded Recrop alongside the normal search. Raw visual confidence selects between the two visual hypotheses.

We selected three representative methods spanning two lines of work related to TGTrack. GlobalTrack and ELGLT extend the search beyond the local crop through full-image search and verifier-triggered re-detection, respectively. PVT++ uses temporal motion prediction to compensate for tracking latency.

AspectGlobalTrackELGLTPVT++TGTrack
ObjectiveFull-image trackingFailure recoveryLatency compensationRecover missing local evidence
Motion modeling------Motion–visual predictionIrregular-time, confidence-aware geometry
Role of prediction------Estimate latency-compensated outputPropose a bounded search region
Search policyFull-image search each frameVerifier-triggered global re-detectionTemporal prediction for latencyExtra Recrop on predicted crop exit
ReliabilityVisual matchingVisual verification---Raw visual confidence across crops
At a glanceSearch the full frameRe-detect globally on demandPredict a latency-corrected statePredict where to look, then select visually

GlobalTrack and ELGLT can search the full image, broadening coverage and introducing more potential distractors for small targets with weak appearance cues. TGTrack limits the extra search to one bounded Recrop.

PVT++ predicts a target state to compensate for tracking latency. TGTrack encodes each observation's age and pairwise time gaps, and weights observations by tracking confidence. The resulting geometry guides when and where to Recrop, and raw visual confidence selects the final box from the two search regions. Removing the observation-age embedding alone reduces TGTrack's AUC by 1.14 points.

Quantitative comparison on AntiUAV410. To test whether existing re-detection or prediction mechanisms provide similar gains in this setting, we adapt ELGLT and PVT++ to LoRAT-L. We integrate ELGLT's recovery controller, verifier, and Guider, fine-tune the auxiliary networks on the AntiUAV410 training split, and use its default recovery thresholds. For PVT++, we train the joint motion–visual predictor and a channel adapter on the same split. At inference, its one-step prediction repositions the single search crop for the next frame and the LoRAT-L visual head produces the box.

MethodBaseline AUCResult AUCChange
GlobalTrack (reported)---65.10---
LoRAT-L + adapted ELGLT63.3462.28-1.06
LoRAT-L + adapted PVT++63.3464.56+1.22
LoRAT-L + TGTrack63.4566.31+2.86

Under the evaluated adaptations, temporal prediction helps when recentering a single crop (+1.22 AUC), while the ELGLT-style recovery mechanism lowers AUC by 1.06 points. TGTrack improves AUC by 2.86 points over its own baseline.

Second, the proposed Temporal Geometry Branch explicitly models sparse and irregular trajectory observations.

During tracking, unreliable visual observations are excluded from trajectory Memory, resulting in irregular temporal intervals. TGTrack addresses this characteristic through age-aware encoding, confidence-weighted temporal aggregation, learned horizon fusion, and adaptive motion correction.

In the original Table 5, under the same visual tracker, conditional Recrop mechanism, and fixed Hann weighting, the proposed Temporal Geometry achieves 65.16% AUC, outperforming Constant Velocity (63.91%) by 1.25 points and Generic Learnable Geometry (64.27%) by 0.89 points.

To further identify the contributions of individual components, we conduct six additional ablations within the full TGTrack framework. Each variant is retrained under the same training protocol, with only the corresponding component removed.

ConfigurationAUC (%)ΔAUC
Full TGTrack66.31--
w/o Age Embedding65.17-1.14
w/o Relative-Age Bias65.79-0.52
w/o Confidence Weighting65.33-0.98
w/o Query Pooling65.37-0.94
w/o Learned Horizon Fusion64.36-1.95
w/o Adaptive Motion Correction65.52-0.79

All six components contribute positively to tracking performance. In particular, removing Learned Horizon Fusion decreases AUC by 1.95 points, highlighting the importance of explicitly conditioning geometric prediction on the elapsed time since the latest admitted observation.

Age Embedding and Confidence Weighting contribute 1.14 and 0.98 AUC points, respectively, further supporting the value of modeling temporal irregularity and observation reliability. These controlled ablations clarify why the proposed geometry branch provides advantages over a generic trajectory Transformer.

Third, the improvement is not merely attributable to more frequent additional searches.

To evaluate search-allocation effectiveness under comparable computational budgets, we conduct supplementary experiments on AntiUAV410. For this evaluation, we independently retrain the learnable geometry models and reevaluate all search-allocation strategies under a common experimental protocol.

All configurations use the same visual tracker, conditional Recrop mechanism, and fixed Hann weighting (λh=0.45\lambda_h=0.45), without Gate–Hann.

For the budget-controlled variants, eligible Recrop requests are temporally subsampled at predetermined rates. A fixed random seed determines the initial sampling phase for each sequence, after which requests are retained according to a deterministic periodic schedule. This procedure does not use ground-truth information, prediction errors, or candidate quality to select requests.

Search AllocationAUCSARecrop (%)Executions
Constant Velocity63.3964.511.83402,376
Generic Learnable Geometry63.6364.762.09202,711
Proposed (Low Budget)64.8966.030.97091,258
Proposed (Near-Matched Budget)64.9766.111.87622,431
Proposed (Full)65.3566.499.609012,450

Under closely matched execution budgets, the proposed Temporal Geometry achieves 64.97% AUC at a 1.8762% Recrop rate, outperforming Constant Velocity (63.39% AUC at 1.8340%) by 1.58 AUC points with only 55 additional executions.

Moreover, it outperforms Generic Learnable Geometry by 1.34 AUC points while executing 280 fewer Recrops (2,431 vs. 2,711). Even at a substantially reduced execution rate of 0.9709%, the proposed method maintains 64.89% AUC, exceeding both simpler strategies.

These results provide further evidence that the gain originates from more effective temporal geometry-guided search allocation, rather than simply allocating more visual searches.

Furthermore, Gate–Hann provides reliability-aware spatial regularization within the allocated Recrop. In additional fixed-weight comparisons, Gate–Hann achieves 66.31% AUC, exceeding the best tested fixed Hann weight (65.98%) by 0.33 points, further supporting the benefit of adapting the spatial prior to predicted localization reliability.

Taken together, the methodological comparison, controlled geometry experiments, component-wise ablations, and budget-controlled evaluations clarify the technical contribution of TGTrack: time- and reliability-aware trajectory modeling enables more effective acquisition of additional visual evidence, while the final tracking hypothesis remains determined by current-frame appearance.

W3. Runtime and computational overhead

On the same AntiUAV410 setup with an RTX 4090, throughput decreases from 349.47 to 292.01 FPS (16.4%). The reciprocal of throughput gives 2.86 versus 3.42 ms per frame, an increase of 0.56 ms per frame on average.

Recrop runs on 6.19% of frames. The table reports its compute cost averaged over all frames.

Model or componentParameters (M)GMACs per frameFPS
LoRAT-L baseline307.43103.02349.47
Temporal Geometry overhead+1.23+0.01---
Conditional Recrop overhead, averaged---+6.38---
TGTrack308.66109.41292.01

The extra visual pass contributes 6.38 of the 6.39 additional GMACs per frame. At this cost, TGTrack improves AUC by 2.86 points on AntiUAV410. A strict real-time deadline depends on per-frame latency, particularly on frames that invoke Recrop.

Many thanks for your thoughtful feedback.

Best regards,
On behalf of the authors

Reviewer_FCxnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TGTrack, a temporal geometry-aware local tracking framework for thermal infrared anti-UAV tracking. It extends LoRAT-L with a trajectory-based geometry branch to predict target location and reliability, enabling conditional local recropping, while Gate–Hann adaptively adjusts spatial priors. Experiments on AntiUAV410 and AntiUAV show AUC improvements of 2.86 and 4.03 points over the controlled LoRAT-L baseline, respectively, with an inference speed of 292.01 FPS.

Critical strengths
  1. Practical and well-motivated design. TGTrack separates search allocation from visual selection, using trajectory geometry to guide additional search while retaining appearance-based candidate selection. This is a reasonable approach to recovering targets outside local search regions.
  2. Consistent experimental improvements. TGTrack improves AUC by 2.86 points on AntiUAV410 and 4.03 points on AntiUAV under direct cross-dataset transfer. The ablation studies also provide evidence for the contributions of the proposed components.
Critical weaknesses
  1. Limited technical novelty. The proposed method mainly combines trajectory-based motion prediction, adaptive local search, and confidence-guided spatial weighting. While the combination is reasonable, the individual ideas are closely related to existing tracking approaches. The paper does not fully establish what makes this combination a substantial technical advance over prior motion-guided tracking methods.
  2. Limited evaluation scope. The experiments are mainly built around LoRAT-L and thermal infrared anti-UAV benchmarks. Although cross-dataset evaluation is included, it remains unclear whether the proposed approach generalizes to other tracking architectures or more diverse tracking scenarios. This limits the evidence for its broader applicability.
Questions for authors

Q1. How does TGTrack differ technically from existing motion-guided tracking and adaptive search-region methods? What specific advantage does the proposed temporal geometry modeling offer beyond combining trajectory prediction with an additional local search?

Q2. The evaluation mainly focuses on LoRAT-L and thermal infrared anti-UAV benchmarks. Can the authors provide evidence that TGTrack can also work with other tracking architectures, or clarify whether its design relies on specific properties of LoRAT-L?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear all,

We sincerely thank all reviewers for their valuable feedback. In response, we have conducted extensive experiments to clarify TGTrack's technical contributions and validate its effectiveness and generalizability.

1. Technical contribution and controlled search allocation.

We clarify how TGTrack differs from existing re-detection and temporal prediction methods: temporal geometry guides where to acquire additional visual evidence, rather than directly determining the final target location. Our adapted baseline comparisons, six component-wise ablations, and budget-controlled experiments further validate this design. Under near-matched Recrop execution rates, our temporal geometry achieves +1.58 AUC over Constant Velocity and +1.34 AUC over Generic Learnable Geometry, supporting more effective search allocation rather than simply increasing search frequency.

2. Geometric prediction, recovery, and reliability-aware adaptation.

We provide additional analyses connecting geometric prediction errors to Recrop coverage, visual candidate quality, and final tracking performance. Across 22 evaluated occlusion events, the recovery rate improves from 40.91% to 72.73%. We also introduce fixed-weight and shuffled-gate controls: Gate–Hann outperforms the best tested fixed Hann weight by +0.33 AUC and Shuffled Gate by +0.35 AUC. Additional gate–error analyses support the role of predicted reliability in adaptive spatial regularization.

3. Broader applicability and efficiency.

We extend evaluation beyond LoRAT-L and thermal infrared anti-UAV benchmarks. TGTrack improves AUC on OSTrack-256 (+1.14) and LETrack (+2.54), as well as on general SOT benchmarks including NfS-30 (+2.78), LaSOT (+0.24), GOT-10k validation (+0.17), and TLPattr (+0.90).

Notably, Recrop is triggered on only 0.66% of frames on LaSOT, 0.014% on GOT-10k validation, and 0.76% on TLPattr, compared with 6.19% on AntiUAV410 in the original TGTrack evaluation. These limited opportunities explain why there are smaller aggregate gains on general SOT benchmarks. Nevertheless, TGTrack achieves positive improvements, including +6.66 AUC on TLPattr's out-of-view subset, supporting its applicability beyond thermal infrared tracking.

Overall, these additional experiments highlight TGTrack's contribution: learning temporal geometry from sparse, irregular, and reliability-varying observations to guide efficient visual search allocation. We will incorporate these results and clarifications into the revised manuscript.

We sincerely appreciate the reviewers' and the Area Chair's time and valuable suggestions.

Best regards,
On behalf of the authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.