acceptodds
Under review as a conference paper at ICLR 2027

UAV-Diverse: A Benchmark for Multi-Person Tracking in Diverse UAV Scenarios

Abstract

Robust multi-person tracking across diverse UAV scenarios is critical for real-world UAV applications. Existing UAV multi-object tracking (MOT) benchmarks are typically developed for specific application scenarios, providing limited insight into the cross-scene transfer capability of current trackers. To address this gap, we present UAV-Diverse, a benchmark for evaluating multi-person tracking across diverse UAV scenes. UAV-Diverse evaluates trackers without additional dataset-specific training, providing a unified protocol for assessing cross-scene transfer across three major scenario groups, including competitive sports, wilderness search and patrol, and urban and indoor public spaces, with nine fine-grained subscenes in total. The benchmark exposes key challenges including identity recovery after target disappearance, drastic viewpoint changes, dynamic camera motion, large target scale variations, and severe appearance ambiguity. Extensive evaluations demonstrate that representative MOT methods suffer substantial performance degradation, highlighting the significant challenges that remain in robust multi-person tracking in diverse UAV scenarios. We further provide a simple training-free baseline to facilitate future research.

Primary area:
datasets and benchmarks
Keywords:
Multi-Object Tracking, UAV Vision, benchmark

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_cBL5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces UAV-Diverse, a test-only benchmark for multi-person tracking in UAV videos. The dataset contains 44 videos, 37,363 frames, 1,498 trajectories, and approximately 821,900 bounding boxes. It covers three major scenario categories (sports, wilderness search-and-rescue/patrol, and urban/indoor environments) and nine sub-scenarios. The benchmark also preserves identity consistency when people leave the field of view and later re-enter.

The authors compare UAV-Diverse with VisDrone and MMOT using normalized entropy and other statistical measures. They argue that UAV-Diverse presents greater challenges in viewpoint variation, camera motion, object scale, appearance, and re-entry events. They also suggest that existing multi-object tracking (MOT) methods have limited cross-scene generalization.

The paper further proposes UAV-Tracker, a training-free baseline built on a frozen SAM3 model. It consists of three components: object lifecycle management, a context-based trajectory stabilizer, and scene fingerprint-based policy adaptation. UAV-Tracker achieves a HOTA score of 61.90 on the proposed benchmark.

Critical strengths

1. A valuable benchmark for UAV-based person tracking.

UAV-Diverse provides a person-centered tracking benchmark with diverse scenarios. It also maintains consistent identity annotations when people leave and re-enter the field of view. This makes the dataset valuable for studying long-term identity preservation and person re-identification. In addition, three annotators who were not involved in bounding-box annotation independently labeled viewpoint, camera motion, and appearance conditions in a blinded manner. This is a well-designed annotation procedure.

2. Detailed experimental analysis.

The appendix provides results grouped by different conditions, including re-entry events, object scale, viewpoint, and camera motion. The authors also report threshold sensitivity in Table 4. Importantly, they acknowledge potential confounding factors between these subsets. These analyses help readers better understand the benchmark's challenges and the limitations of the reported results.

Critical weaknesses

1. The main claims about limited cross-scene generalization and benchmark difficulty lack sufficient evidence. Some results even contradict these claims.

  • Detector-related confounding factors. The methods in the Original Pipeline group use detectors trained on VisDrone (HOTA 26–35), while those in the Shared SAM3 group use a foundation model detector (HOTA 52–62). This large gap may mainly reflect differences in detection quality rather than cross-scene tracking ability. Since the benchmark has no in-domain training set, it is also difficult to separate domain shift from the inherent difficulty of the dataset.

  • UAV-Diverse appears easier than VisDrone under the same detector. With SAM3 detection, ByteTrack achieves 54.56 HOTA on UAV-Diverse but only 33.15 on the VisDrone pedestrian subset. Similarly, UAV-Tracker achieves 61.90 and 37.47, respectively (Table 2 vs. Table 10). Although differences in dataset composition may affect this comparison, these results weaken the claim that UAV-Diverse is more challenging.

  • The proposed statistics do not directly measure difficulty. Normalized entropy measures how evenly category labels are distributed, not how difficult the tracking task is. The reported 97.5% conditional re-entry rate largely reflects the annotation protocol. UAV-Diverse reuses the original identity when a person re-enters, while datasets such as VisDrone may assign a new identity. Moreover, only 7.9% of trajectories (118 trajectories) contain valid disappearance events.

2. The baseline offers limited improvements, and the evaluation lacks important comparisons for identity association.

  • UAV-Tracker relies heavily on hand-crafted rules. Equations 15–23 introduce many fixed coefficients and thresholds, which the authors tune on five rugby videos. This makes the method relatively complex for a "simple baseline" and raises concerns about its general applicability.

  • The performance gains are modest. Compared with AMOT+SAM3Feat, UAV-Tracker improves HOTA by only 2.11 points. However, its MOTA is lower (58.20 vs. 62.83). The authors also report results from only a single run, without variance estimates.

  • The ReID comparisons are insufficient. For identity recovery, UAV-Tracker only uses cosine similarity between SAM3 features. The authors do not compare it with trackers that use dedicated ReID models, such as StrongSORT, Deep OC-SORT, or BoT-SORT with OSNet/CLIP-ReID features. Therefore, it is unclear whether the improvement comes from the proposed lifecycle management or the quality of the appearance features.

  • Potential annotation bias. The benchmark uses SAM2 for pre-annotation, while the proposed tracker relies on SAM3. Any remaining annotation bias may favor SAM-based methods. The paper does not sufficiently examine this possibility.

  • Dataset availability and reproducibility. Some videos come from online sources, and it is unclear whether the authors have permission to redistribute the original footage. This may affect the reproducibility and long-term availability of the benchmark.

Questions for authors
  1. Under the same SAM3 detector, tracking scores on UAV-Diverse are much higher than those on the VisDrone pedestrian subset. How do the authors justify the claim that UAV-Diverse is more challenging? Could the authors provide an in-domain training upper bound, for example using cross-validation splits, to separate the effects of domain shift from dataset difficulty?

  2. How do VisDrone and MMOT handle identity annotations when a person leaves and re-enters the field of view? How much of the difference between 97.5% and 26.1% comes from annotation protocols rather than actual scene characteristics?

  3. Could the authors include trackers with dedicated ReID features as additional baselines? It would also be useful to replace the SAM3 features in UAV-Tracker with ReID features to evaluate their respective contributions.

  4. What are the licensing conditions for the videos collected from online sources? If the original videos cannot be redistributed, how can other researchers access the complete test dataset?

  5. Could the authors report results across multiple runs or different hyperparameter settings to assess performance variance? A sensitivity analysis of UAV-Tracker's scene fingerprint coefficients would also help evaluate its robustness.

5 comments
AuthorsOpenReview ↗

Response to Reviewer cBL5

We appreciate the reviewer's careful assessment and recognition of UAV-Diverse's diverse scenario coverage, identity-consistent annotations, and detailed challenge analyses. We also value the constructive suggestions regarding ReID comparisons, robustness, and dataset reproducibility. UAV-Diverse is the primary contribution of our paper. Below, we provide clarifications and additional experiments to address these concerns.

1. Evaluation Protocol and Cross-Dataset Performance (W1.1, W1.2, Q1)

Cross-dataset performance and benchmark scope.

As documented in Appendix A.5.3 and Figure 6, we identified annotation issues in VisDrone, including oversized ground-truth boxes and unannotated visible people. These issues can affect human-only tracking scores. We recognize the importance of reliable annotations for cross-dataset comparisons. However, comprehensively reannotating VisDrone would require a separate dataset curation effort beyond the scope of this work.

We would like to clarify that UAV-Diverse does not claim to be inherently harder than VisDrone. Our focus is evaluating cross-scene transfer across diverse UAV environments. The reported statistics characterize challenge coverage and diversity, not absolute tracking difficulty.

Evaluation protocols.

Section 5.1 defines two complementary protocols. Original Pipeline evaluates five VisDrone-trained trackers using their native inference pipelines. Shared SAM3 Detections controls detection inputs to compare tracking strategies.

UAV-Diverse is intentionally test-only. Its purpose is to evaluate cross-scene transfer under domain shift, without target-domain training or adaptation. In-domain cross-validation would instead evaluate dataset-specific adaptation under a different protocol.

2. Diversity Statistics and Identity Reappearance (W1.3, Q2)

Diversity versus difficulty.

We would like to clarify the distinction between challenge diversity and tracking difficulty. Section 3.2 uses normalized entropy to characterize challenge diversity. UAV-Diverse achieves 89.0% viewpoint entropy (VisDrone: 28.6%, MMOT: 0.0%), 91.0% camera-motion entropy (74.4%, 39.6%), and 89.6% appearance entropy (14.3%, 15.6%). Its target-scale range spans 14.9×, compared with 7.3× and 3.6×.

These results demonstrate broader coverage of heterogeneous UAV conditions. Appendix A.3 further examines tracking performance under these challenges. For example, in the smallest-target regime, AMOT+SAM3Feat achieves only 18.35 HOTA, compared with 38.33 for UAV-Tracker (Appendix Table 6), highlighting the challenges of small-target tracking and the differences between tracking methods.

Identity consistency and reappearance.

We would like to clarify that the reported 97.5% is a conditional reappearance rate. Although valid disappearance events involve 7.9% of trajectories, reappearance events occur in 24 of the 44 sequences. Under the same 15-frame criterion, 115 of 118 trajectories with valid disappearance events satisfy the reappearance criterion in UAV-Diverse (97.5%), compared with 71/272 (26.1%) in VisDrone and 32/54 (59.3%) in MMOT. Both VisDrone-MOT and MMOT explicitly specify temporally consistent identity annotations in their annotation protocols. Appendix Table 4 confirms that the relative ordering remains unchanged across 10-, 15-, and 30-frame thresholds.

AuthorsOpenReview ↗

Response to Reviewer cBL5

3. Baseline Design, Performance Gains, and ReID Comparisons (W2.1–W2.3, Q3)

Identity management and mechanism contributions.

UAV-Tracker improves identity management and trajectory continuity on top of frozen SAM3. Table 3 evaluates lifecycle management (OLM) under identical SAM3 observations.

ConfigurationHOTA ↑AssA ↑IDF1 ↑
Native SAM351.6461.1661.05
SAM3 + OLM59.4868.9568.06
Improvement7.84 (+15.2%)7.79 (+12.7%)7.01(+11.5%)

The full UAV-Tracker further achieves 61.90 HOTA, a 10.26-point (+19.9%) improvement over Native SAM3. To assess whether simpler mechanisms suffice, we evaluate simplified OLM reactivation, a fixed CS propagation budget, and disabled SF under identical SAM3 observations.

ConfigurationMechanism changedHOTA ↑AssA ↑IDF1 ↑MOTA ↑
Simplified OLM ReactivationCosine threshold only (0.85)61.167.7170.6358.15
Fixed CS BudgetFixed propagation budget B=561.6269.271.8258.09
Without SFSF disabled61.3569.0771.1756.2
Full UAV-TrackerOriginal configuration61.969.4371.9758.2

The full method improves HOTA by 0.28–0.80 points over the simplified alternatives. As the reviewer notes, all parameters were selected using five external rugby videos and applied unchanged to all 44 UAV-Diverse sequences (Appendix A.6), supporting cross-scene applicability without target-domain tuning.

Dedicated ReID comparisons.

We appreciate the reviewer's helpful suggestion to strengthen the ReID comparison. Table 2 includes BoT-SORT with a VisDrone-trained FastReID SBS-S50 model and ReID enabled. Following the reviewer's suggestion, we further replace UAV-Tracker's SAM3 appearance features with the same frozen FastReID model while keeping its other mechanisms unchanged.

MethodAppearanceHOTA ↑AssA ↑IDF1 ↑
BoT-SORTFastReID54.5160.1463.64
UAV-TrackerSAM361.969.4371.97
UAV-TrackerFastReID62.0269.7872.19

With the same FastReID model and shared SAM3 detections, UAV-Tracker outperforms BoT-SORT by 7.51 HOTA (+13.8%), 9.64 AssA (+16.0%), and 8.55 IDF1 (+13.4%). Replacing SAM3 appearance features with FastReID also slightly improves UAV-Tracker's HOTA from 61.90 to 62.02. These results, together with the OLM ablation, support the contribution of the proposed mechanisms beyond the choice of appearance features.

Performance gains and evaluation metrics.

Compared with AMOT+SAM3Feat, UAV-Tracker improves HOTA by 2.11, AssA by 3.29, and IDF1 by 3.17 points, while achieving lower MOTA (58.20 vs. 62.83). We acknowledge this trade-off. As discussed by Luiten et al. (IJCV 2021, Sec. 9.1), MOTA is less informative about identity association, whereas HOTA balances detection and association accuracy. Given the importance of identity continuity in our benchmark, we emphasize HOTA, AssA, and IDF1 while also reporting MOTA for a comprehensive evaluation.

4. Evaluation Stability and Annotation Reliability (W2.1, W2.4, Q5)

Evaluation stability.

As detailed in Appendix A.6, all inference parameters were fixed across evaluation datasets. OLM, SF, and adaptive propagation make deterministic tracking decisions without random sampling given identical SAM3 observations and inference settings. A full-dataset rerun reproduced the archived predictions across all 44 UAV-Diverse sequences, with identical track IDs and bounding-box coordinates in the evaluation coordinate system.

We further evaluate Scene Fingerprint coefficient sensitivity under four configurations, independently varying the coverage and identity-risk weights by ±20% while retaining scene-conditioned adaptation. Across all 44 sequences, HOTA remains between 61.8982 and 61.8987, indicating stable performance within the tested parameter range.

Annotation Quality and Potential SAM2 Pre-annotation Bias.

We recognized the importance of reliable identity annotations. SAM2 was used only to initialize annotations, rather than to define the final ground truth. As detailed in Appendix A.1, all bounding boxes and identity labels were manually reviewed and refined in CVAT, with ambiguous cases resolved through additional review and consensus. All trackers are evaluated against the same manually verified ground truth.

AuthorsOpenReview ↗

Response to Reviewer cBL5

5. Dataset Availability and Licensing (W2.5, Q4)

We appreciate the reviewer's concern about dataset accessibility, licensing, and reproducibility. For online-source videos, we will adopt a URL-based release policy, following established practices in TrackVerse (ICCV 2025), PreVAD (ICLR 2026), and VALUE (NeurIPS 2021). We will not host or redistribute third-party video files or extracted frames. Instead, we will release source URLs, precise timestamps, frame mappings, MOT annotations, and evaluation code, allowing researchers to reconstruct the test clips from the original sources where access remains available and permitted.

The original videos remain subject to their owners' rights and applicable licensing terms, while our annotations and code will have separate licenses. We have updated the Ethics and Reproducibility Statements in the revised PDF to clarify this release policy and privacy safeguards.

We hope these results and clarifications help address the reviewer's concerns. Thank you again for your time and consideration.

Reviewer_w9ufOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UAV-Diverse, a benchmark for evaluating the cross-scene transferability of multi-object tracking (MOT) of people across diverse UAV (unmanned aerial vehicle) scenarios. The benchmark covers three main groups (nine sub-scenes in total), namely competitive sports, wilderness search and patrol, and urban and indoor public spaces, and it systematically evaluates key challenges such as identity recovery after target disappearance and viewpoint changes. The paper also proposes UAV-Tracker, a training-free baseline built on a frozen SAM3 model, and shows that its adaptive tracking policy based on scene fingerprints achieves higher tracking performance than existing methods, particularly in identity preservation.

Critical strengths
  • Previous UAV-MOT benchmarks were limited to specific domains such as traffic or seaside scenes. This work integrates diverse scenarios into a framework for evaluating cross-scene transferability, which is highly valuable for practical UAV applications.
  • Rather than merely reporting metrics, the paper provides a quantitative analysis along five dimensions, including viewpoint, camera motion, target scale, and appearance ambiguity. This analysis clearly identifies where current MOT methods suffer performance degradation.
  • By proposing UAV-Tracker, which leverages the powerful SAM3 foundation model to achieve training-free adaptive tracking, the paper establishes a fair point of comparison for future research.
Critical weaknesses
  • Although the dataset is designed for diversity, the number of 44 videos is small for deep-learning-based evaluation. Validation on a larger dataset would be desirable to sufficiently verify the proposed method's generalization performance.
  • Although the work assumes deployment on UAVs, it lacks a detailed analysis of inference speed (FPS) and computational resource requirements. Because it relies on a large model such as SAM3, its feasibility on actual hardware remains questionable.
  • In the "Shared SAM3 Detections" comparison group, the detector is unified across methods. This may bias the evaluation in favor of the proposed method over traditional MOT methods, where detector optimization plays an important role.
  • The text in the figures (particularly Figure 2) is too small and should be enlarged to a size comparable to the main text.
Questions for authors
  • Can the authors provide specific figures for the inference speed and memory consumption of UAV-Tracker? In particular, I would like to hear the authors’ view on its feasibility on embedded devices mounted on UAVs.
  • How was the size of the causal window (30 frames) in the Scene Fingerprint determined? Is there any data showing how performance changes as this window size varies?
  • The paper notes that the annotations in VisDrone-MOT are incomplete. If a more precise, person-only subset were used, do the authors expect UAV-Tracker to retain its advantage?
4 comments
AuthorsOpenReview ↗

Response to Reviewer w9uf

We appreciate the reviewer's recognition of UAV-Diverse's cross-scene evaluation value, five-dimensional challenge analysis, and training-free baseline. We address the remaining concerns below.

1. Dataset Scope and Evaluation Fairness (W1, W3)

Dataset scope.

We appreciate the reviewer's suggestion to further expand the evaluation dataset. To provide additional context, we compare the test-set scale with existing UAV-MOT benchmarks.

BenchmarkTest SequencesAvg. Duration
VisDrone-MOT201916—
MMOT5022.1s
UAV-Diverse (Ours)4428.1s

UAV-Diverse contains 37.4K test frames, compared with 6.3K for VisDrone-MOT2019 and approximately 5.4K for MMOT. We agree that expanding the dataset would further strengthen the evaluation and consider how we can use a more powerful automated pipeline to scale the dataset size in our future work.

Evaluation fairness.

We acknowledge the reviewer's concern that shared SAM3 detections may favor UAV-Tracker. As explained in Section 5.1 and Table 2, we adopt two complementary protocols to accommodate different tracking architectures. The Original Pipeline group preserves each method's native inference pipeline, as replacing the detector in end-to-end, query-based, or detector-coupled methods would alter their original designs. For methods that support external detections, the Shared SAM3 Detections group uses identical detection inputs to compare tracking strategies under controlled conditions.

2. Computational Efficiency and Deployment (W2, Q1)

We thank the reviewer for the constructive suggestion regarding computational efficiency and practical deployment. We evaluate UAV-Tracker on an NVIDIA RTX 3090 across all 44 sequences.

MethodFPS ↑Peak GPU Memory ↓
Native SAM31.018.96 GiB
UAV-Tracker0.7410.12 GiB

FPS is measured over all frames. Memory reports the median per-sequence peak GPU allocation.

As acknowledged in our Limitations section, UAV-Tracker inherits the computational cost of SAM3. Its current inference speed is insufficient for real-time onboard deployment. As UAV-Tracker serves as a baseline, we encourage the community to improve inference efficiency while maintaining tracking performance in the future.

3. Causal Window Sensitivity (Q2)

We understand the reviewer's suggestion to examine the sensitivity to the causal window length. Scene Fingerprint uses a default window of 30 frames, corresponding to approximately one second of tracking history at 30 FPS. We therefore evaluate window lengths of 15, 30, and 60 frames across all 44 sequences, keeping other settings unchanged.

WindowHOTA ↑AssA ↑IDF1 ↑
15 frames61.899269.431771.976
30 frames (default)61.898469.433971.9738
60 frames61.896569.428471.9706

HOTA varies by only 0.0027 points, with similarly small variations in AssA and IDF1. These results demonstrate stable tracking performance across different causal window lengths, supporting the robustness of UAV-Tracker to this design choice.

4. VisDrone Evaluation and Figure Readability (Q3, W4)

VisDrone evaluation.

We appreciate the reviewer's constructive comments on the VisDrone evaluation. To examine this more directly, we identified two VisDrone test sequences captured in different environments, both containing confirmed annotation errors. We manually corrected their annotations and reevaluated the trackers. The HOTA results are reported below.

SequenceMethodOriginal GTCorrected GT
uav0000119_02301_vAMOT26.8931.75
UAV-Tracker36.6942.18
uav0000249_02688_vAMOT26.1516.38
UAV-Tracker12.7532.66

Under the corrected annotations, UAV-Tracker achieves higher HOTA than AMOT on both sequences.

Figure readability.

We appreciate the reviewer's suggestion regarding Figure 2. We have enlarged the text and improved label readability in the revised PDF.

We hope these results and clarifications help address the reviewer's concerns. Thank you again for your time and consideration.

We thank the reviewer again for the constructive feedback. We have provided detailed responses and additional experimental evidence to address the concerns raised.

We would appreciate hearing whether our responses have addressed your main concerns, or if any questions remain. We would be happy to provide further clarification.

In the authors' response, appropriate supplements and additional experiments have been provided to address the concerns I raised. Based on the responses addressing dataset scale, fairness of the evaluation protocol, computational efficiency, and causal window sensitivity, I believe that the usefulness of the dataset, the validity of the evaluation methods, and the reliability of the proposed method are clearly demonstrated.

AuthorsOpenReview ↗

Thank you for your time and comments

Dear Reviewer w9uf,

Thank you very much for your positive feedback and thoughtful consideration of our response. We sincerely appreciate your constructive suggestions, which have helped us strengthen our paper.

Best regards,
Authors of Paper #5661

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

This paper introduces UAV-Diverse, a benchmark for evaluating cross-scene generalization in UAV-based multi-person tracking. The authors evaluate existing MOT methods under a unified protocol without dataset-specific training and propose UAV-Tracker, a training-free baseline built on frozen SAM3. On shared SAM3 detections, UAV-Tracker achieves 61.90 HOTA, outperforming AMOT+SAM3Feat. The benchmark addresses a relevant problem, but its distinctiveness relative to existing datasets, annotation reliability, and the justification of the baseline's rule-based design require further validation.

Critical strengths
  1. The paper addresses cross-scene generalization in UAV-based human tracking, covering sports, wilderness, urban, and indoor environments. Its emphasis on identity recovery after disappearance is particularly useful for evaluating long-term tracking robustness.
  2. Evaluating trackers with shared SAM3 detections helps isolate differences in tracking strategies. UAV-Tracker improves HOTA, AssA, and IDF1 over AMOT+SAM3Feat, and the component ablations provide initial evidence for the proposed design.
Critical weaknesses
  1. The Object Lifecycle Manager, Scene Fingerprint, and Context-Guided Track Stabilizer rely on manually designed lifecycle transitions, diagnostic signals, and context-dependent control rules. Although the ablations show that these components improve aggregate tracking metrics, they do not establish why the specific mechanisms are preferable to simpler alternatives, such as fixed thresholds, simpler reactivation policies, or direct use of tracking diagnostics without scene-conditioned adaptation.
  2. The benchmark uses sequence-level identities and retains identity labels across target reappearance, making annotation consistency particularly important for AssA and IDF1. The described manual refinement and quality-control pipeline is useful, but the paper does not report quantitative inter-annotator agreement or identity-consistency/error statistics on a reviewed subset. In addition, the condition-specific analysis considers factors such as camera motion, scale variation, and viewpoint, which can co-occur within the same sequences.
Questions for authors
  1. How were the key thresholds and policy-update rules selected without using UAV-Diverse annotations?
  2. Have simpler alternatives or sensitivity analyses been evaluated to establish the contribution of Scene Fingerprint and the proposed gating mechanisms?
  3. What specific evaluation capabilities does UAV-Diverse provide beyond existing datasets, particularly TeamTrack, VisDrone, and MMOT? How many sequences overlap with TeamTrack, and how does this affect the claimed cross-dataset evaluation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer EgAV

We thank the reviewer for recognizing UAV-Diverse's diverse scenario coverage, its value for long-term identity recovery, and the controlled evaluation of tracking strategies. We address the remaining concerns with additional evidence below.

1. Method Design and Parameter Selection (W1, Q1, Q2)

Parameter selection.

As detailed in Appendix A.6, the key thresholds were selected using five external rugby videos disjoint from UAV-Diverse. The policy-update rules were also fixed before evaluation and applied uniformly across all sequences, without using UAV-Diverse annotations.

Mechanism comparisons.

Following the reviewer's suggestion, we evaluate simplified OLM reactivation and a fixed CS propagation budget. We also include the Without SF ablation, which retains OLM and CS while disabling scene-conditioned policy adaptation. All comparisons use the same frozen SAM3 observations.

ConfigurationMechanism changedHOTA ↑AssA ↑IDF1 ↑MOTA ↑
Simplified OLM ReactivationCosine threshold only (0.85)61.167.7170.6358.15
Fixed CS BudgetFixed propagation budget B=561.6269.271.8258.09
Without SFSF disabled61.3569.0771.1756.2
Full UAV-TrackerOriginal configuration61.969.4371.9758.2

The full method achieves higher scores than all tested alternatives. On the 24 reappearance sequences, OLM recovery checks also reduce identity switches from 816 to 763 compared with simplified reactivation.

These comparisons also help clarify the design choices in Section 4. A fixed propagation budget cannot adjust the permitted duration of SAM3-only hypotheses to changing tracking conditions, while cosine-only reactivation lacks the temporal and layout checks used to validate identity recovery. Accordingly, CS constrains propagation using recent detector support, OLM verifies identity recovery through multiple consistency checks, and SF adapts their shared policy based on past tracking diagnostics.

2. Annotation Reliability and Challenge Analysis (W2)

Annotation reliability.

We appreciate the reviewer raising the question of annotation consistency. As described in Appendix A.1, all bounding boxes and identities were manually refined and verified in CVAT. Ambiguous cases were reviewed by an additional annotator and resolved through consensus. To further assess annotation reliability, an independent reviewer examined all 115 valid reappearance trajectories across 24 sequences and found no identity-assignment errors (0/115).

Overlapping challenge conditions.

We understand the reviewer's concern regarding overlapping challenge conditions. Since UAV-Diverse consists of real-world UAV footage, multiple challenges inevitably co-occur within individual sequences. However, we would like to point out that, as explicitly stated in Lines 1011–1013, these results should not be interpreted as isolated causal effects of individual challenge dimensions. In addition, all methods are evaluated on the same sequence subsets under the same evaluation protocol, ensuring fair comparisons across methods despite overlapping conditions.

3. Benchmark Distinctiveness and TeamTrack Overlap (Q3)

Benchmark evaluation capabilities.

TeamTrack evaluates full-pitch tracking across multiple sports, VisDrone supports general multi-category UAV tracking, and MMOT focuses on multispectral UAV tracking with oriented annotations. UAV-Diverse complements these benchmarks by providing a unified, test-only evaluation of cross-scene transfer across sports, wilderness search and patrol, and urban/indoor environments. It further supports long-term identity recovery evaluation with valid reappearance events (Appendix Table 5), and condition-specific analysis of viewpoint, camera motion, target scale, and appearance ambiguity (Appendix A.3).

Relationship with TeamTrack.

For the source video, 8 of the 44 sequences are sourced from TeamTrack's test split, within the sports category. However, the original TeamTrack annotations primarily cover athletes and the ball. For UAV-Diverse, we removed ball annotations, refined nearly all existing human bounding boxes to our full-body standard, and added bounding boxes and identities for previously unlabeled visible people. All annotations underwent the same identity verification and quality-control protocol (Appendix A.1). The reported evaluation therefore includes eight TeamTrack-derived sports sequences and 36 sequences from other sources, covering sports, wilderness, and urban/indoor environments.

We hope these results and clarifications help address the reviewer's concerns. Thank you again for your time and consideration.

We thank the reviewer again for the constructive feedback. We have provided detailed responses and additional experimental evidence to address the concerns raised.

We would appreciate hearing whether our responses have addressed your main concerns, or if any questions remain. We would be happy to provide further clarification.

Dear Reviewer EgAV,

We sincerely thank you again for the time and effort you have devoted to reviewing our work.

Following our earlier responses, we have provided additional experimental results and clarifications to address your concerns.

As the discussion period draws to a close, we would like to kindly follow up and ask whether our responses have adequately addressed your concerns. We would be happy to provide any further clarification if needed. If you find that your concerns have been resolved, we would greatly appreciate it if you could consider revisiting your assessment.

Best regards,

Authors of Paper #5661

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.