acceptodds
Under review as a conference paper at ICLR 2027

MSR-TRACK: MITIGATING SEARCH REPRESENTATION DILUTION IN INFRARED ANTI-UAV TRACKING

Abstract

Reliable infrared anti-UAV tracking is difficult because aerial targets occupy few pixels, provide weak appearance cues, and are frequently obscured by clutter, thermal distractors, rapid motion, or temporary disappearance. Although adaptive local trackers enlarge the search area to recover targets, resizing the larger crop to a fixed input shrinks the target’s relative scale and admits additional background. We call this effect search representation dilution. To address it, we introduce MSR-Track, which enhances search representations at complementary token and spatial stages. Lightweight Token Refinement (LTR) recalibrates search tokens before template–search interaction without altering their number or spatial order. The Multi-Scale Response Enhancement (MSRE) head aggregates fine-scale evidence and broad context to reinforce weak UAV responses before bounding-box prediction. MSR-Track obtains 62.5% AUC, 86.1% Precision, and 82.2% Normalized Precision on Anti-UAV410, exceeding our FocusTrack-SRA by 2.3, 2.9, and 2.7 points while adding little computation. Tests on Anti-UAV300 and CST Anti-UAV demonstrate cross-dataset generalization under unseen distributions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
visual object tracking, infrared small target tracking, vision transformers, token refinement, multi-scale representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GHuiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the degradation of target representations caused by enlarging the search region in infrared anti-UAV tracking. The authors refer to this issue as search representation dilution, where resizing an enlarged search crop to a fixed input resolution reduces the relative size of the target while introducing additional background information.

To address this problem, the authors propose MSR-Track, which builds upon FocusTrack-SRA and introduces two modules: Lightweight Token Refinement (LTR), which aggregates and redistributes search-token information before template–search interaction, and Multi-Scale Response Enhancement (MSRE), which combines spatial features at multiple scales before bounding-box prediction.

On Anti-UAV410, MSR-Track achieves 62.5% AUC, outperforming the reproduced FocusTrack-SRA baseline by 2.3 points, with relatively small computational overhead. Additional experiments include component ablations, parameter sensitivity, attribute-wise evaluations, and cross-dataset tests on Anti-UAV300 and CST Anti-UAV.

Critical strengths

1、The paper addresses an important issue in infrared anti-UAV tracking, where enlarging the search region improves target coverage but also reduces the relative target resolution and introduces additional background clutter. The proposed modules preserve the existing search policy, input resolution, and prediction interface, making the approach relatively easy to integrate into an established tracking framework.

2、The experimental evaluation covers multiple datasets, component ablations, attribute-wise performance, and computational efficiency. Both LTR and MSRE provide measurable improvements over the reproduced baseline, and their combination achieves the best results. The reported computational overhead is also modest, suggesting practical value for efficient infrared tracking.

Critical weaknesses

1、 The authors explicitly claim that LTR can reduce background-dominant interference before template–search interaction. However, LTR operates only on search tokens (x) without access to the target template (z), and the paper does not adequately explain how it distinguishes target-relevant information from visually similar distractors without target-specific guidance. Meanwhile, the baseline already incorporates global self-attention and a CLS token, yet the paper lacks sufficient justification or controlled comparisons to establish why additional latent aggregation and gated redistribution are necessary. Similarly, the authors claim that the coarse-scale branches of MSRE help suppress structured distractors. However, average pooling and multi-scale feature fusion primarily provide broader spatial context, which does not inherently demonstrate how structured background interference is suppressed. Although the component ablations, attribute-wise evaluations, and qualitative results demonstrate performance improvements, they do not directly verify the claimed interference suppression mechanisms or sufficiently isolate the contributions of these specific design choices from general feature enhancement.

2、The conceptual novelty of search representation dilution is not sufficiently established. The described phenomenon largely follows from the expected consequences of enlarging a search crop while maintaining a fixed input resolution: the target occupies a smaller relative area, and more background clutter is introduced. Although the paper provides a scaling analysis and a search-factor study in Figure 1, it remains unclear what additional insight this formulation offers beyond the known trade-off between search coverage and representation quality.

Questions for authors

1、Given that LTR processes only search tokens before template–search interaction, how does it distinguish target-relevant information from visually similar distractors without access to the target template? Could the authors provide further evidence supporting the claimed interference suppression mechanism and clarify why latent aggregation and gated redistribution are necessary beyond the baseline's existing global self-attention and CLS token?

2、The authors claim that the coarse-scale branches of MSRE help suppress structured distractors. However, broader contextual aggregation does not necessarily imply distractor suppression. Could the authors provide more direct evidence, such as response-map visualizations or distractor-specific analyses, to support this explanation?

3、Beyond the expected reduction in relative target size and increase in background clutter caused by enlarging the search crop, what additional insight does the proposed concept of search representation dilution provide? Could the authors further clarify how this phenomenon differs from the known trade-off between search coverage and representation quality?

4、Given the unresolved 2.1-point AUC discrepancy in the reproduced FocusTrack-SRA baseline, could the authors provide additional verification or evidence supporting the robustness of the reported 2.3-point improvement?

5、Could the authors clarify the detailed training configurations for both stages, including training duration, learning rates, and frozen or trainable components, to facilitate reproducibility and verify the stated consistency of the training protocols?

7 comments
AuthorsOpenReview ↗

Response to Reviewer GHui – Q1

Q1. Given that LTR processes only search tokens before template–search interaction, how does it distinguish target-relevant information from visually similar distractors without access to the target template? Could the authors provide further evidence supporting the claimed interference suppression mechanism and clarify why latent aggregation and gated redistribution are necessary beyond the baseline’s existing global self-attention and CLS token?

Response:

We sincerely thank the reviewer for this important question. We agree that, because LTR receives only search tokens before template–search interaction, it cannot identify the true target or suppress an appearance-similar distractor according to target identity at this stage. Target identity is determined by the subsequent template–search interaction. We will therefore describe LTR more precisely as template-independent search-representation refinement rather than as a module that identifies the target before seeing the template.

(1) Latent-token aggregation. Enlarging the adaptive search region introduces more background tokens into a fixed-size input grid, which can dilute the representation of a tiny target. LTR uses eight latent tokens to form multiple complementary summaries of the complete search region. These summaries serve only as internal information carriers: they neither replace nor remove the original 256 search tokens. Consequently, the full dense spatial grid remains available for subsequent template-conditioned matching and localization.

(2) Position-dependent gated write-back. Uniformly broadcasting a single global summary would provide the same supplementary information to every candidate location. Instead, LTR selects a different combination of latent summaries for each search position and writes it back residually. The gates do not determine the target in advance; they condition each candidate representation on its intra-search context, after which the template–search interaction performs target identification.

(3) Relation to global self-attention and the CLS token. Global self-attention provides general token-to-token contextualization, whereas LTR explicitly implements a multi-summary aggregation and position-dependent write-back path. A CLS token represents a single global state and does not explicitly assign different combinations of multiple summaries to individual dense positions. We do not claim that LTR is universally “more necessary” than either mechanism. Rather, it provides a complementary structural bias for background-dominated dense search representations.

(4) Fixed-input evidence. To eliminate trajectory-induced crop differences, we provided all models with pixel-identical template and search crops. The results were:

MetricMean change (×100)95% sequence-clustered bootstrap CI (×100)
Target-response gain+1.35[+0.25, +2.66]
Fixed non-target reduction+0.59[−0.15, +1.54]
Target–non-target margin gain+1.94[+0.39, +3.84]

These values are raw [0,1] response-score changes multiplied by 100, not percentages or AUC points. The exploratory analysis used 50 fixed frames from 21 sequences. The target-response and margin-gain intervals were above zero, whereas the interval for fixed non-target reduction included zero. Thus, the evidence supports stronger target evidence and improved target–non-target separation, but not uniform suppression of every distractor.

Accordingly, we will replace the claim that “LTR suppresses distractors” with the more precise statement that “LTR reorganizes the search representation and improves target–non-target separability after template-conditioned interaction.”

AuthorsOpenReview ↗

Response to Reviewer GHui – Q2

Q2. The authors claim that the coarse-scale branches of MSRE help suppress structured distractors. However, broader contextual aggregation does not necessarily imply distractor suppression. Could the authors provide more direct evidence, such as response-map visualizations or distractor-specific analyses, to support this explanation?

Response:

We sincerely thank the reviewer for identifying the important distinction between broader contextual aggregation and distractor suppression. We agree that the presence of multiscale branches alone cannot establish that non-target responses are reduced. The comment prompted us to refine both the mechanistic interpretation of MSRE and the scope of our claim.

(1) Function of the multiscale branches. MSRE retains a fine spatial branch and several coarser contextual branches. The fine branch preserves local details required for tiny-target localization, while the coarse branches capture structural relationships over larger spatial extents. Their outputs are aligned to the same dense grid, fused, and written back residually to the original representation. Thus, MSRE directly enables the prediction head to use local responses together with broader context; it is not an explicit distractor-removal operator.

(2) Why broader context may still help. Strong heat sources, edges, and repeated textures can produce local infrared responses similar to those of a tiny target, but their spatial extent and surrounding structure may differ over a larger neighborhood. Multiscale features allow the prediction head to consider both a local peak and its contextual structure, reducing reliance on an isolated response. This is a plausible design motivation, not evidence that every structured distractor is consistently suppressed.

(3) Fixed-input response evidence. To remove trajectory-induced crop differences, all compared variants received pixel-identical template and search crops. The fixed non-target region was selected from the SRA response map and evaluated at the same location for every variant. We compared complete MSRE with SRA and also disabled the coarse branches of the same trained MSRE checkpoint at inference time:

MetricComplete MSRE vs. SRA, mean [95% CI] (×100)Coarse branches ON vs. OFF, mean [95% CI] (×100)
Target-response gain+1.18 [−0.57, +3.38]+1.15 [+0.37, +1.94]
Fixed non-target reduction+1.13 [+0.46, +2.07]−0.20 [−0.50, +0.05]
Target–non-target margin gain+2.30 [+0.11, +5.05]+0.94 [+0.03, +1.82]

The values are raw [0,1] response-score changes multiplied by 100, not percentages or AUC points. The exploratory analysis used 50 fixed frames from 21 sequences. Coarse ON vs. OFF was an inference-time diagnosis, not a separately retrained ablation. Complete MSRE reduced the selected non-target response and improved the response margin on average. For the coarse branches, the stable effects were target-response enhancement and margin improvement; their non-target reduction was not stable. Therefore, the results do not support the claim that every coarse branch directly suppresses distractors.

(4) Evidence boundary and revised claim. Demonstrating suppression of specific categories, such as heat sources, building edges, or cloud textures, would require category-level annotations and measurements at pixel-identical locations. Our current experiments did not include such semantic annotations, so we will not interpret general non-target regions as validated examples of specific physical distractors.

Accordingly, we will replace “the multiscale branches of MSRE suppress structured distractors” with: “MSRE combines local details with multiscale context to enhance the search representation and improve target–non-target separability. This design may help with spatially structured high-response backgrounds, but the current evidence does not support universal suppression across distractor categories.”

AuthorsOpenReview ↗

Response to Reviewer GHui – Q3

Q3. Beyond the expected reduction in relative target size and increase in background clutter caused by enlarging the search crop, what additional insight does the proposed concept of search representation dilution provide? Could the authors further clarify how this phenomenon differs from the known trade-off between search coverage and representation quality?

Response:

We thank the reviewer for asking how search representation dilution relates to the established trade-off between search coverage and representation quality. We agree that resizing a larger search crop to a fixed input resolution reduces the relative size of the target and brings more background into the input. FocusTrack also discusses feature degradation as the search factor increases. We use the term search representation dilution to describe the representation-quality side of this known trade-off in tiny infrared target tracking. Our focus is on how the tracker uses the limited target information available at a fixed input resolution once the target is within the search region.

Figure 1 of the original submission provided additional empirical evidence. We evaluated FocusTrack-SRA and MSR-Track at search factors of 4, 5, and 6 while keeping the input resolution fixed. Both trackers achieved higher AUC values as the search factor increased. Meanwhile, the AUC gain of MSR-Track over FocusTrack-SRA increased from 1.6 percentage points at a search factor of 4 to 3.0 percentage points at a search factor of 6. Under the tested settings, both widening the search region and enhancing the search representation improved tracking performance, and MSR-Track achieved a larger gain over the baseline at wider search regions.

Changing the search factor affects target coverage, resized target size, and background content together. We therefore interpret Figure 1 as a performance trend observed under their joint variation. In the revised manuscript, we will clarify how the term relates to the established trade-off and frame the contribution as an empirical analysis of the gains from search feature enhancement across search factors.

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

The paper argues that enlarging the search region in adaptive local anti-UAV trackers, then resizing it to a fixed input, shrinks already-tiny infrared targets and lets in more background, an effect it calls “search representation dilution.” On top of FocusTrack-SRA, it adds LTR, which pools search tokens into a few latent summaries and gates them back into every token before the backbone, and MSRE, a pyramid-pooling block applied to the final search map before the box head. MSR-Track reaches 62.5% AUC on Anti-UAV410, 2.3 points above the authors’ reproduction of FocusTrack-SRA at nearly the same cost, with ablations, a search-factor study, and direct transfer without fine-tuning to Anti-UAV300 and CST Anti-UAV as further evidence.

Critical strengths

The motivation is intuitive and real: widening the search crop shrinks an already-tiny infrared target, and Figure 1 supports this by showing the gain growing from 1.6 to 3.0 AUC as the search factor increases. The design is lean and deployment-friendly, preserving all interfaces and staying above 130 FPS at only about 0.7% extra MACs.

Critical weaknesses

LTR is essentially TokenLearner with its TokenFuser write-back, and MSRE closely follows PSPNet's Pyramid Pooling Module, which is not cited. The contribution is therefore a combination of known components inside one tracker, offering little new insight at the level ICLR expects.

The reproduced baseline scores 2.1 AUC below its published result, nearly matching the claimed 2.3-point improvement, and MSR-Track exceeds the published FocusTrack-SRA number by only 0.2.

Questions for authors

How do LTR and MSRE differ in substance from TokenLearner/TokenFuser and PSPNet’s Pyramid Pooling Module? An ablation that swaps in these modules directly would show whether your design changes matter.

What explains the 2.1-point gap between your reproduced FocusTrack-SRA and the published result? Do LTR and MSRE still give a clear gain when added to the official code or checkpoint?

3 comments
AuthorsOpenReview ↗

Response to Reviewer FMpq – Q1

Q1. How do LTR and MSRE differ in substance from TokenLearner/TokenFuser and PSPNet’s Pyramid Pooling Module? An ablation that swaps in these modules directly would show whether your design changes matter.

Response:

We thank the reviewer for pointing out the connections between LTR and TokenLearner/TokenFuser, and between MSRE and the Pyramid Pooling Module (PPM) in PSPNet. We agree that direct substitution experiments would help determine whether our design choices offer additional benefits over the corresponding existing modules. Our experiments evaluated the gains relative to a shared tracking baseline but did not include these direct substitutions. Therefore, a broader claim of superiority over TokenLearner/TokenFuser or PPM would require the comparisons suggested by the reviewer. In the revised manuscript, we will add the relevant citations, including PSPNet, explain these relationships explicitly, and clarify the scope of our architectural contributions and performance claims.

Our study focuses on search representation dilution in adaptive local tracking. We build on existing ideas in token learning and pyramid pooling and adapt them to the spatial requirements of dense tracking.

(1) Differences between LTR and TokenLearner/TokenFuser. In the TokenLearner/TokenFuser pipeline, dense features are first compressed into a small set of tokens. Subsequent processing operates on these compact tokens, and TokenFuser then maps the resulting features back to the spatial grid. LTR uses eight latent summaries internally to aggregate information across the search region. Before template–search interaction, position-dependent gates write information from these summaries back to the original 256 search tokens. The subsequent Transformer therefore performs template–search interaction over the full search grid. LTR uses compact summaries to provide each search location with global context from the search region while preserving the identity and spatial order of all search positions. This preserves the dense spatial correspondence needed to localize tiny targets.

(2) Differences between MSRE and PSPNet’s Pyramid Pooling Module. Both MSRE and PPM use multiscale pooling to gather context. They differ in how they use the pooled features and how they connect to the task head. PSPNet’s PPM is designed for semantic segmentation. It uses several fixed output grids to collect contextual information and concatenates the resulting context features with the original feature map. MSRE operates directly on the final 16×16 search feature map. It uses one branch at the original resolution and three pooled branches with output grids of 8×8, 4×4, and 2×2. After aligning the branch outputs to a common resolution, MSRE fuses them and adds the fused features back to the original search feature map through a learnable residual scale. The resulting features feed into the existing bounding-box prediction head. This design is intended to preserve tiny-target responses and incorporate background context at several spatial scales within the dense localization grid and the existing prediction interface. These architectural differences explain how we adapted multiscale pooling to the tracking task. The direct substitutions proposed by the reviewer would test the performance implications of those choices.

(3) Updated ablation results and scope of the conclusions. Under the updated common training configuration, FocusTrack-SRA achieved 62.3% AUC. Adding LTR or MSRE individually increased the AUC to 63.1% and 63.3%, respectively, while using both modules yielded 63.7%. These results quantified the individual gains of the two modules over the same tracking baseline and showed that their combination achieved a higher AUC than either module alone. We will therefore describe LTR and MSRE as effective architectural adaptations for dense infrared anti-UAV tracking and relate their design to search representation dilution. Our performance conclusions will remain limited to the measured gains over the shared baseline. We appreciate the reviewer’s suggestions regarding direct module substitution and comparisons at matched computational cost, and we plan to pursue these analyses in future work.

AuthorsOpenReview ↗

Response to Reviewer FMpq – Q2

Q2. What explains the 2.1-point gap between your reproduced FocusTrack-SRA and the published result? Do LTR and MSRE still give a clear gain when added to the official code or checkpoint?

Response:

We thank the reviewer for pointing out the reproduction gap and for suggesting that we reassess LTR and MSRE on a baseline that matches the published AUC.

In the original submission, our independently reproduced FocusTrack-SRA achieved 60.2% AUC on Anti-UAV410, 2.1 percentage points below the published SRA-only result. To investigate this gap, we re-examined the training configuration, particularly the global batch size. Our initial single-GPU configuration used a global batch size of 64. In the updated experiments, we increased the global batch size to 128 on a single NVIDIA A40 and retrained FocusTrack with SRA enabled and ATM disabled. The retrained baseline achieved 62.3% AUC, matching the published SRA-only result to the reported decimal place.

We subsequently reran all four ablation variants under the same updated training configuration. Relative to the retrained baseline, adding LTR alone increased the AUC to 63.1% (+0.8 percentage points), while adding MSRE alone increased it to 63.3% (+1.0 percentage point). The model with both modules achieved the highest AUC, 63.7%. It also improved precision from 86.2% to 87.5% and normalized precision from 82.3% to 83.7%.

Table R1. Anti-UAV410 ablations under the original and updated training configurations

ModelOriginal AUC (%), batch size 64Updated AUC (%), batch size 128Updated gain (pp)
FocusTrack-SRA60.262.3—
+ LTR61.763.1+0.8
+ MSRE61.963.3+1.0
+ LTR + MSRE62.563.7+1.4

The updated gains were calculated relative to the retrained FocusTrack-SRA baseline under the same batch-size-128 protocol.

The 2.3-percentage-point gain reported in the original submission was obtained under the earlier training configuration. Under the updated configuration, both modules again improved tracking individually, and their combination yielded a 1.4-percentage-point AUC gain over the retrained baseline, whose performance matched the published AUC.

Dear Reviewer,

The authors have responded to your review. Could you please take a moment to review their responses and engage in the discussion?

Thank you for your time and valuable contributions.

Kind regards, AC

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

This paper studies the expansion of search regions in infrared anti-UAV tracking. Enlarging the local search region can improve target coverage, but at a fixed input resolution it also reduces the target's relative size and introduces more background. The authors call this effect "search representation dilution" and add two modules to FocusTrack-SRA. Before template-search feature interaction, LTR aggregates search information through a small set of latent tokens and writes it back to the search tokens. Before prediction, MSRE applies multi-scale pooling and residual fusion to the final search feature map.

The paper reports 62.5% AUC on Anti-UAV410, a gain of 2.3 percentage points over the authors' reproduction of FocusTrack-SRA. Computation increases from 29.1 G MACs to 29.3 G, while speed decreases from 139.3 FPS to 131.9 FPS. The authors also provide module ablations, experiments with fixed search factors, and direct transfer results on Anti-UAV300 and CST Anti-UAV.

Critical strengths
  1. Tracking accuracy improves at a small computational cost. In Table 6, the added modules increase MACs by only about 0.7% and retain about 94.7% of FocusTrack-SRA's processing speed. If the results can be reproduced consistently, this improvement has practical value and should be easy to incorporate into existing local trackers.

  2. Both modules improve performance individually and in combination. In Table 3, LTR and MSRE alone achieve 61.7% and 61.9% AUC, respectively, while their combination reaches 62.5%. Table 2 also shows improvements when testing directly on two other datasets, indicating that the gains occur across multiple test sets.

Critical weaknesses
  1. The reproduced FocusTrack-SRA used for comparison performs below the original result, for an unexplained reason. Section 4.1 reports 60.2% AUC for the reproduction, compared with 62.3% in the original paper. This gap of 2.1 percentage points is close to the 2.3-point improvement reported here. The original paper also reports 62.8% AUC for the full FocusTrack model. The current results show that the proposed method outperforms the authors' reproduced version, but it remains unclear whether adding the two modules would yield the same improvement when FocusTrack-SRA reaches its original performance. Scores across papers may depend on training and evaluation settings, so the authors need to explain the gap and repeat the comparison under the same conditions.

  2. The experiments do not yet explain why these two modules are preferable to simple changes. LTR aggregates latent tokens and writes the information back through gates, while MSRE pools, projects, and fuses multi-scale features. The ablations show gains after adding the modules, but do not establish whether ordinary modules with a similar number of added parameters could achieve similar results. Figure 1 shows larger gains with larger search regions. However, expanding the search region also changes whether the target is included in the crop, its relative size, and the amount of background. The effects of these factors are not separated. Comparisons with simple modules of similar computational cost or parameter count, together with analyses of target and background responses on identical crops, are needed to show whether the improvement specifically addresses "representation dilution."

Questions for authors
  1. Could the authors use the official FocusTrack weights and evaluation procedure to identify why their reproduction is 2.1 percentage points below the original result? What are the results of adding LTR and MSRE when FocusTrack-SRA reaches its original performance?

  2. Which parameters are updated in each training stage, how many epochs does each stage use, and what are the optimizer and learning rate? Are the total numbers of epochs and updates comparable across FocusTrack-SRA and the ablation variants? Is the 2.3-point improvement stable across multiple independent training runs?

  3. Do LTR and MSRE still offer an advantage over residual MLPs, ordinary spatial convolutions, or pooling-and-fusion modules with similar computational cost? On identical search crops, do the added modules make target and background responses easier to distinguish?

4 comments
AuthorsOpenReview ↗

Response to Reviewer baYV – Q1

Q1. Could the authors use the official FocusTrack weights and evaluation procedure to identify why their reproduction is 2.1 percentage points below the original result? What are the results of adding LTR and MSRE when FocusTrack-SRA reaches its original performance?

Response:

We thank the reviewer for pointing out the reproduction gap and for suggesting that we reassess LTR and MSRE on a baseline that matches the published AUC.

In the original submission, our independently reproduced FocusTrack-SRA achieved 60.2% AUC on Anti-UAV410, 2.1 percentage points below the published SRA-only result. To investigate this gap, we re-examined the training configuration, particularly the global batch size. Our initial single-GPU configuration used a global batch size of 64. In the updated experiments, we increased the global batch size to 128 on a single NVIDIA A40 and retrained FocusTrack with SRA enabled and ATM disabled. The retrained baseline achieved 62.3% AUC, matching the published SRA-only result to the reported decimal place.

We subsequently reran all four ablation variants under the same updated training configuration. Relative to the retrained baseline, adding LTR alone increased the AUC to 63.1% (+0.8 percentage points), while adding MSRE alone increased it to 63.3% (+1.0 percentage point). The model with both modules achieved the highest AUC, 63.7%. It also improved precision from 86.2% to 87.5% and normalized precision from 82.3% to 83.7%.

Table R1. Anti-UAV410 ablations under the original and updated training configurations

ModelOriginal AUC (%), batch size 64Updated AUC (%), batch size 128Updated gain (pp)
FocusTrack-SRA60.262.3—
+ LTR61.763.1+0.8
+ MSRE61.963.3+1.0
+ LTR + MSRE62.563.7+1.4

The updated gains were calculated relative to the retrained FocusTrack-SRA baseline under the same batch-size-128 protocol.

The 2.3-percentage-point gain reported in the original submission was obtained under the earlier training configuration. Under the updated configuration, both modules again improved tracking individually, and their combination yielded a 1.4-percentage-point AUC gain over the retrained baseline, whose performance matched the published AUC.

AuthorsOpenReview ↗

Response to Reviewer baYV – Q2

Q2. Which parameters are updated in each training stage, how many epochs does each stage use, and what are the optimizer and learning rate? Are the total numbers of epochs and updates comparable across FocusTrack-SRA and the ablation variants? Is the 2.3-point improvement stable across multiple independent training runs?

Response:

We thank the reviewer for asking about the training setup and the consistency of the results. In the updated experiments, we matched the total training budget across FocusTrack-SRA, the LTR-only and MSRE-only variants, and the full model. The full model was trained using the two-stage procedure described in Section 3.4 of the original submission.

(1) Training stages, trainable parameters, and optimization

All four variants were initialized from the same DropMAE-pretrained ViT-B/16 backbone weights. We trained FocusTrack-SRA and the two single-module variants for 40 epochs each. For FocusTrack-SRA, we jointly updated the Transformer backbone, the SRA head, and the bounding-box prediction head. The LTR-only and MSRE-only variants additionally updated their respective modules.

We trained the full model in two stages. In Stage 1, we trained the MSRE-enhanced tracker for 30 epochs and jointly updated the Transformer backbone, MSRE, the SRA head, and the bounding-box prediction head. In Stage 2, we loaded the checkpoint from epoch 30, kept MSRE enabled, inserted LTR, and optimized only the LTR parameters for 10 epochs. The Transformer backbone, MSRE, the SRA head, and the bounding-box prediction head were frozen during Stage 2, although they remained active in the forward pass and gradients propagated through them to LTR. The full model therefore received 40 epochs of training in total, with LTR optimized during the final 10 epochs.

In Stage 1, we used AdamW with a base learning rate of 4e-4 and a weight decay of 1e-4. The Transformer backbone used a learning-rate multiplier of 0.1, corresponding to a learning rate of 4e-5, while MSRE and the prediction heads used the base learning rate. In Stage 2, we used AdamW with an initial learning rate of 5e-5 for LTR and the same weight decay of 1e-4. We applied cosine annealing over the 10 Stage-2 epochs and set the gradient-clipping threshold to 0.1.

(2) Matched training budgets for the baseline and ablation variants

All four variants used the same training data, sampling and augmentation strategies, loss functions, and global batch size of 128. We sampled 85,715 training examples per epoch. With drop_last=True in the data loader, single-GPU training yielded 669 optimizer updates per epoch. FocusTrack-SRA and each single-module variant received 26,760 optimizer updates over 40 epochs. The full model received 20,070 updates in Stage 1 and 6,690 in Stage 2, resulting in the same total of 26,760 updates. Thus, all four variants were matched in terms of total training epochs and optimizer updates.

(3) Consistency across the two training configurations

With the original batch size of 64, FocusTrack-SRA, the LTR-only variant, the MSRE-only variant, and the full model achieved AUCs of 60.2%, 61.7%, 61.9%, and 62.5%, respectively. With the updated batch size of 128 and a different GPU setup, the corresponding AUCs were 62.3%, 63.1%, 63.3%, and 63.7%.

In both sets of independently trained runs, LTR and MSRE each improved on the baseline, MSRE achieved slightly higher AUC than LTR, and the combination achieved the highest AUC. The full model’s gain was 2.3 percentage points in the original setup and 1.4 percentage points in the updated setup. The direction of the gains, the relative ranking of the single-module variants, and the benefit of combining them were consistent across the two training configurations and GPU environments. The revised manuscript will report the updated results obtained under the matched total training budget as the main results.

AuthorsOpenReview ↗

Response to Reviewer baYV – Q3

Q3. Do LTR and MSRE still offer an advantage over residual MLPs, ordinary spatial convolutions, or pooling-and-fusion modules with similar computational cost? On identical search crops, do the added modules make target and background responses easier to distinguish?

Response:

We thank the reviewer for suggesting comparisons with alternatives of similar computational cost and a response analysis on identical search crops.

(1) Comparison with alternatives of similar computational cost

Before template–search interaction, LTR aggregates information across search locations and writes it back to the original search tokens through position-dependent gates. Before the prediction head, MSRE combines fine-grained spatial features with multiscale context. A token-wise residual MLP, a standard spatial convolution, and a simple pooling-and-fusion block are appropriate controls for testing these design choices.

In the updated experiments, we matched the total training budget across all four variants. FocusTrack-SRA achieved 62.3% AUC on Anti-UAV410. Adding LTR or MSRE individually raised the AUC to 63.1% and 63.3%, respectively, and the full model reached 63.7%. These ablations quantified the gains over the shared baseline. Determining whether LTR and MSRE outperform the suggested alternatives would require direct substitutions at the same insertion points with comparable additional computational cost.

(2) Target and non-target responses on identical template and search crops

To control for crop differences caused by tracking trajectories, we performed a fixed-input analysis using the earlier checkpoints trained with a global batch size of 64. We selected 50 frames from 21 sequences and gave all compared variants pixel-identical template and search crops. For each frame, we selected a non-target region using the FocusTrack-SRA response map and evaluated the same region for all variants. The table summarizes the response-margin changes relative to FocusTrack-SRA.

ComparisonMean increase in target–non-target response margin (×100)Sequence-clustered 95% bootstrap CI (×100)
LTR-only vs. FocusTrack-SRA+1.94[0.39, 3.84]
MSRE-only vs. FocusTrack-SRA+2.30[0.11, 5.05]

The response margin is the target response minus the response in the selected non-target region. The table reports changes in this margin on the raw response-score scale multiplied by 100. Both single-module variants increased the mean response margin on the evaluated frames. For LTR, the confidence interval for the increase in target response was above zero. For MSRE, the confidence interval for the reduction in non-target response was above zero.

Dear Reviewer,

The authors have responded to your review. Could you please take a moment to review their responses and engage in the discussion?

Thank you for your time and valuable contributions.

Kind regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.