acceptodds
Under review as a conference paper at ICLR 2027

U-Revisit: Persistent Event Reasoning under Intermittent Aerial Revisits

Abstract

Route-based UAV inspection efficiently covers spatially distributed regions but observes each location only intermittently, with irregular gaps during which the region state remains unobserved. Verifying persistent events therefore requires reasoning across revisits while distinguishing confirmed absence from missing or unreliable evidence. A region-centered benchmark and a coarse-to-fine pipeline are introduced for this setting. Fixed semantic map regions provide spatial correspondence across revisits, and a lightweight proposal stage identifies candidate sequences for a learned multi-revisit verifier. U-Revisit combines visual, telemetry, map, and observation-quality information through quality-aware fusion. Separate spatial and instance memories represent region state and vehicle-related evidence. Under a seven-fold, site-ID-disjoint LOSO protocol, U-Revisit achieves the highest mean Event AP of across three seeds. Single Revisit and Unordered Set achieve and , respectively. Repeated-observation aggregation accounts for most of the Event AP improvement over Single Revisit, and U-Revisit further improves Event AP beyond Unordered Set.

Primary area:
datasets and benchmarks
Keywords:
UAV perception Persistent, event detection, Aerial revisit, Multi-revisit learning, Region-centered benchmark

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ygiJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies persistent event verification under intermittent UAV revisits, e.g., whether a parking or roadway region remains occupied across passes. It introduces a benchmark of 11 sites that organizes observations around fixed semantic map regions and distinguishes confirmed absence from missing or unreliable evidence. The proposed U²-Revisit fuses visual, telemetry, and map features through quality-aware masked fusion, and maintains separate spatial and instance memories with time decay. Under a seven-fold LOSO protocol, it achieves an Event AP of 0.9594.

Critical strengths
  1. The task is practically relevant, and explicitly separating confirmed absence from missing or unreliable observations is a reasonable design for intermittent sensing.
  2. The benchmark uses a site-disjoint LOSO protocol and reports results over multiple seeds.
Critical weaknesses
  1. Insufficient and unconvincing experiments. All baselines are self-constructed variants; no published methods for temporal or irregularly sampled sequence modeling are compared, including those discussed in the related work. Moreover, Unordered Set, which ignores revisit order and time, outperforms U²-Revisit on AUROC, F1, and BAcc, which contradicts the claimed benefit of ordered temporal modeling. The AP gains over the baselines and in the ablations are within the standard deviations over three seeds, so the improvements are not significant.
  2. Core claim is not validated and key results are missing. Distinguishing missing evidence from confirmed absence is presented as the key challenge, but no experiment evaluates it, e.g., robustness on sequences with many unavailable observations. The external-site evaluation on R10/R11, the auxiliary task results, and a description of the "Statistics" baseline (whose AUROC of 0.42 is below chance) are not provided.
  3. Unclear writing and task definition. In particular, the second paragraph of the introduction abruptly describes a "lightweight proposal stage" before the task, the challenges, or the model have been introduced, making it unclear what is being proposed and why. More generally, the introduction lists several difficulties without explaining how each is addressed, and the term "persistent event" is not defined until Section 3.3, where it is still unclear what condition makes an occupancy "persistent."
Questions for authors
  1. How does U²-Revisit compare with published methods for irregular or missing-data sequence modeling? Why does Unordered Set outperform U²-Revisit on AUROC, F1, and BAcc?
  2. Are the update mask and quality descriptors computed automatically, or derived from the annotated evidence states?
  3. What is the "Statistics" baseline, and why is its AUROC below 0.5?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ygiJ (1)

We sincerely thank the reviewer for the detailed and constructive feedback, particularly the concerns regarding experimental validity, missing-evidence robustness, and task clarity. We have substantially expanded our evaluation and revised the manuscript to address these concerns.

CW1 & Q1: Published Baselines and Temporal Modeling

Concern: The original baselines are self-constructed, and the benefits of explicit temporal modeling are not sufficiently established.

Response: We have added four published methods for irregular or incomplete sequence modeling. All methods are evaluated on the same seven-site LOSO benchmark with 544 labeled candidates and three training seeds.

Published baselineYearEvent AP
GRU-D20180.8082
SeFT20200.9357
ATENet20250.7375
MTM20250.9262
U²-Revisit—0.9594

U²-Revisit achieves higher mean Event AP than all four task-adapted published methods. These methods use candidate-level supervision, whereas U²-Revisit additionally uses auxiliary multitask supervision; the comparison therefore evaluates complete task-adapted systems rather than isolating architectural differences.

Regarding Internal Baselines and Capacity Control

We report the strongest internal baselines separately, because they test different questions from the published-method comparison above.

Internal baselineEvent APAUROCF1BAcc
Unordered Set0.94620.94650.93100.9059
Matched GRU (H259)0.95960.94400.91750.8846
U²-Revisit0.95940.94530.92690.9013

Unordered Set. Multi-revisit aggregation itself provides substantial improvements over single-revisit prediction. U²-Revisit improves the primary Event AP metric by 0.0132 (0.9594 vs. 0.9462), while Unordered Set achieves slightly higher AUROC, F1, and balanced accuracy. We report this as a competitive comparison rather than a claim of superiority on every metric.

Parameter-matched GRU. To control for model capacity, we evaluated a single-memory GRU with 1,003,676 parameters, compared with 1,004,175 for U²-Revisit (a difference below 0.05%). The matched GRU obtains 0.9596 Event AP versus 0.9594 for U²-Revisit, while U²-Revisit achieves higher AUROC (0.9453 vs. 0.9440), F1 (0.9269 vs. 0.9175), and BAcc (0.9013 vs. 0.8846). This shows that simple aggregation is a strong capacity-matched baseline; we therefore position the dual-memory design as a task-motivated inductive bias rather than the source of the main Event-AP gain.

Paired bootstrap. The mean AP difference between U²-Revisit and Unordered Set is +0.0132, with a candidate-level 95% CI of [-0.0013, 0.0264] and a site-cluster 95% CI of [-0.0056, 0.0460]. The interval includes zero, which is why we frame the AP comparison as competitive evidence rather than a statistical superiority claim. The stronger evidence for the proposed design comes from the controlled missing-evidence experiments in CW2.

Manuscript Revision: Section 6.3, Table 4, and Appendices D.6–D.7.


AuthorsOpenReview ↗

Response to Reviewer ygiJ (2)

CW2: Missing-Evidence Robustness

Concern: The paper emphasizes distinguishing unavailable evidence from confirmed absence but originally lacks direct experimental validation.

Response: We have added a controlled synthetic missing-evidence experiment using frozen trained models.

At selected revisits, dynamic evidence and associated quality/availability inputs are masked. We compare two strategies under identical missing positions and original time intervals:

  • Skip Update: Preserve elapsed-time decay without incorporating unavailable observations into recurrent memory.
  • Forced Update: Perform recurrent updates at the same masked positions despite the absence of new dynamic evidence.

Additional results (Appendix D.8):

MissingnessSkip Update APForced Update APΔ\DeltaAP
10%0.95610.9522+0.0039
30%0.95370.9336+0.0200
50%0.94890.8926+0.0563

The advantage of skipping updates increases with missingness, reaching +0.0563 Event AP at 50% missingness.

Both strategies retain elapsed-time decay, so this experiment evaluates the recurrent update policy rather than isolating the effect of time decay. Complementarily, the ablation in Table 3 shows that removing time decay decreases mean Event AP from 0.9594 to 0.9521.

We also evaluate 95 sequences containing naturally occurring detector-missing observations, identified from automatic high-recall records. U$^2$-Revisit achieves Event AP 0.7700, compared with 0.7563 for Unordered Set.

These experiments provide complementary evidence for selective state updates, elapsed-time modeling, and performance under incomplete observations.

Manuscript Revision: Sections 5.2–5.4, Table 3, and Appendix D.8, which now includes a quantitative table and an Event AP trend figure.


Q2: Quality Descriptors and Update-Mask Sources

Question: Are the update mask and quality descriptors computed automatically or derived from annotated evidence states?

Response: Both the quality descriptors and temporal update masks are generated through automatic preprocessing, without using manually annotated persistent-event or vehicle-identity labels as model inputs.

The five quality descriptors capture detection confidence, visual-feature availability, image-to-map registration reliability, GPS-field availability, and image sharpness estimated using Laplacian variance. These signals support quality-aware feature fusion.

The temporal update mask is derived from automatically generated observation-state information and controls whether a revisit triggers a recurrent state update. When an update is suppressed, elapsed-time memory decay remains active.

Importantly, manual event and identity annotations are used only as supervision targets, rather than as inputs to the quality-aware fusion or temporal update mechanism. We have also clarified the provenance of the automatically generated quality features and update masks in the implementation details.

Manuscript Revision: Sections 3.3 and 5.2–5.3.


AuthorsOpenReview ↗

Response to Reviewer ygiJ (3)

CW2: External-Site Evaluation and Auxiliary Tasks

Concern: The external-site evaluation and auxiliary-task results are insufficiently documented.

Response: Appendix D.5 reports external-site evaluation. Models trained on the seven development sites are evaluated on held-out R10 and R11, comprising 211 labeled candidates.

MethodEvent APAUROC
Unordered Set0.9630.789
Single Memory0.9720.812
U$^2$-Revisit0.9730.817

U$^2$-Revisit achieves the highest pooled Event AP and AUROC among the original baselines on these sites.

We evaluated the frozen auxiliary heads at positions with valid supervision. The occupancy and revisit-level persistence heads achieve AP values of 0.9948 and 0.9173, respectively.

Identity-related supervision is considerably sparser: the sample-level identity evaluation contains only three replacement-positive sequences, limiting conclusions about replacement recognition. We report the auxiliary results separately from the primary candidate-level Event AP.


Q3: Statistics Baseline and AUROC Below 0.5

Question: What is the Statistics baseline, and why does it achieve AUROC below 0.5?

Response: Statistics is a supervised logistic-regression baseline using nine handcrafted sequence-level features summarizing vehicle detections, detection continuity, empty and unavailable observations, and revisit intervals.

Features are standardized with StandardScaler, followed by class-balanced logistic regression trained independently on the six training sites of each LOSO fold.

Our additional analysis identifies cross-site score misranking as the main explanation for its low pooled AUROC.

EvaluationAUROC
Within-site pairs0.9639
Cross-site pairs0.3061
Site-macro0.8917
Pooled0.4222

The classifier ranks examples relatively well within individual sites, but its scores are poorly aligned across sites. In particular, high-scoring negative candidates from R1 contribute substantially to incorrect cross-site rankings.

Thus, the pooled AUROC below 0.5 does not indicate reversed class labels; it reflects poor cross-site score comparability for this statistical baseline.

Manuscript Revision: We clarified the baseline definition and its pooled-AUROC interpretation in the experimental analysis.


CW3: Task Definition and Presentation

Concern: The introduction presents the proposal stage before defining the task, and the meaning of persistent events remains insufficiently clear.

Response: We have clarified the operational meaning of persistent occupancy in our region-centered setting.

A persistent event refers to an occupancy episode supported by temporally separated observations of the same fixed semantic region, rather than an isolated vehicle detection within a single revisit. The essential distinction is that persistence is assessed from evidence accumulated across revisits, with spatial correspondence established through the reference map.

For example, when a region is observed as occupied during multiple revisits, an intervening unavailable observation does not automatically invalidate the persistence hypothesis. In contrast, a confirmed-empty observation provides evidence of an occupancy interruption and must be distinguished from missing evidence.

Importantly, persistent occupancy is a region-level property, not a vehicle-identity claim. Different vehicles may occupy the same region across revisits, and the UAV cannot directly observe every transition during the intervening gaps. Accordingly, our labels represent evidence-supported persistent occupancy rather than guaranteed uninterrupted occupancy throughout unobserved intervals.

Operationally, rule-based proposals identify sequences with repeated vehicle observations, within-revisit stationary evidence, and spatial consistency. These criteria select candidates for manual review, not positive event labels. Final persistence labels are assigned through review of the available multi-revisit evidence, while vehicle-identity annotations are handled separately.

We have revised the task presentation to distinguish persistent occupancy, confirmed absence, unavailable evidence, and vehicle-identity continuity.

Manuscript Revision: Introduction, Section 3.3, and Appendices A.4–A.5.


Closing Remarks

We thank the reviewer again for the constructive suggestions. The revised manuscript now includes four published baseline comparisons, a parameter-matched architecture comparison, paired statistical analyses, controlled missing-evidence experiments, and clearer task and implementation descriptions.

We believe these additions substantially strengthen the empirical evaluation and clarify the contributions of U$^2$-Revisit. We welcome further discussion and would appreciate the reviewer's reconsideration in light of the additional evidence.

We hope that our responses have addressed your concerns. If any questions or concerns remain, we would be happy to provide further clarifications and engage in further discussion. We sincerely appreciate your time and valuable feedback.

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

This paper introduces U²-Revisit, a framework for persistent-event verification under intermittent UAV observations. The authors construct a region-centered aerial revisit benchmark that distinguishes confirmed absence from missing or unreliable evidence. The proposed framework combines visual, telemetry, and map features through quality-aware multimodal fusion and employs separate spatial and instance memories with elapsed-time decay to model region occupancy and vehicle-related information. Experiments under a seven-site leave-one-site-out protocol show that U²-Revisit achieves an Event AP of 0.9594, outperforming several single-revisit and temporal aggregation baselines.

Critical strengths

The paper addresses persistent-event verification under intermittent UAV observations, where irregular revisit intervals and unreliable sensing make temporal reasoning challenging. Distinguishing confirmed absence from missing evidence is meaningful for aerial surveillance and long-term monitoring.

Critical weaknesses
  1. The proposed framework mainly combines established techniques, including quality-aware gated fusion, time-decayed GRUs, and separate recurrent memories. While these components are reasonable for the task, the paper does not sufficiently demonstrate a fundamentally new mechanism for persistent-event reasoning beyond conventional temporal aggregation and multimodal modeling. The distinction between spatial and instance memories is conceptually motivated, but its technical contribution remains incremental.

  2. Most of the performance gain comes from aggregating repeated observations rather than the proposed architecture. Unordered Set already achieves 0.9462 Event AP, compared with 0.9594 for U²-Revisit. Moreover, Unordered Set achieves slightly higher AUROC and better F1 and balanced accuracy than the full model. Removing the instance memory decreases Event AP by only 0.0045, raising questions about the necessity of the dual-memory design. The paper lacks sufficient evidence that these relatively small gains justify the additional architectural complexity.

  3. Although the paper adopts a site-disjoint evaluation protocol, performance varies considerably across sites. In Appendix D.4, Unordered Set outperforms U²-Revisit on R1 and R2, while several evaluation sites contain fewer than 25 sequences. The external-site results also show only marginal improvements over simpler baselines. More comprehensive analysis is needed to establish the robustness and statistical significance of the claimed generalization advantages.

  4. The benchmark focuses mainly on persistent vehicle occupancy in roadway and parking regions. Furthermore, the proposal stage relies on repeated detections and spatial consistency, while sequences without valid persistence labels are excluded from evaluation. It remains unclear how well the resulting benchmark represents more challenging events involving prolonged occlusion, missed detections, or intermittent occupancy changes. The effect of proposal-stage recall on overall event verification performance is also insufficiently evaluated.

Questions for authors
  1. What fundamentally distinguishes the proposed dual-memory architecture from conventional multi-branch recurrent models? Can the authors provide more direct comparisons with parameter-matched GRUs, temporal Transformers, or alternative memory architectures to establish the necessity of the proposed design?

  2. Given the relatively small Event AP improvement over Unordered Set and Single Memory, can the authors provide paired statistical significance tests and more detailed analysis of scenarios where the dual-memory design offers clear advantages?

  3. How robust is the proposed framework under varying revisit intervals, registration errors, missing observations, and detector failures? Can the authors provide controlled evaluations to demonstrate that the quality-aware fusion and time-decay mechanisms specifically address these challenges?

  4. What is the recall of the initial event proposal stage, and how does it affect end-to-end persistent-event detection? Can the authors clarify whether the benchmark includes sufficiently challenging negative cases and events missed by the initial proposal rules?

4 comments
AuthorsOpenReview ↗

Response to Reviewer JJQh (1)

We thank the reviewer for the rigorous assessment of our methodological novelty, baseline comparisons, cross-site robustness, and candidate-proposal protocol. These concerns prompted us to examine more carefully which improvements can be attributed to multi-revisit aggregation, temporal modeling, and the dual-memory design.

We address the four main concerns and corresponding questions below.

CW1: Novelty Beyond Existing Temporal Modeling Components

Concern: Quality-gated fusion, time-decayed recurrent updates, and dual-memory processing are individually established techniques. Their combination may represent an incremental architectural contribution.

Response: The central contribution of U$^2$-Revisit lies in the region-centered formulation of persistent-event reasoning under intermittent aerial revisits, together with a quality-aware temporal framework designed for this setting.

Unlike continuous video tracking, this task requires reasoning over observations of fixed semantic regions separated by irregular time intervals, where missing evidence must be distinguished from confirmed absence. Our formulation explicitly considers three complementary aspects: region occupancy, vehicle-identity evidence, and observation reliability.

To address these challenges, U$^2$-Revisit combines spatial memory for region-level event reasoning, instance memory for identity-related auxiliary learning, elapsed-time decay for irregular revisit intervals, and quality-aware fusion for incomplete observations.

The expanded evaluation further supports these design choices. Ablation studies demonstrate the contributions of quality-aware fusion and time decay, while controlled missing-evidence experiments show an improvement of 0.0563 Event AP from selective recurrent updates over forced updates at 50% missingness.

Together, the proposed benchmark, task-oriented temporal formulation, and additional experimental evidence establish the contribution of U$^2$-Revisit to persistent-event reasoning under intermittent aerial observations.


CW2 & Q1: Strong Unordered Baseline and Parameter-Matched Comparisons

Concern: Unordered Set already performs strongly, the performance margins are modest, and the ablation of instance memory shows only a small reduction. Would simpler or parameter-matched alternatives perform similarly?

Response: We expanded the baseline evaluation to include a parameter-matched single-memory GRU and four task-adapted published sequence-modeling methods.

The main results on the seven-site LOSO benchmark are:

MethodEvent APAUROCF1BAcc
GRU-D0.80820.77460.79980.7156
SeFT0.93570.92940.79420.6295
ATENet0.73750.69160.75270.6099
MTM0.92620.92060.89720.8568
Unordered Set0.94620.94650.93100.9059
Matched GRU (H259)0.95960.94400.91750.8846
U$^2$-Revisit0.95940.94530.92690.9013

Comparison with Unordered Set. U$^2$-Revisit achieves a mean Event AP of 0.9594, improving upon Unordered Set (0.9462) by 0.0132. Unordered Set remains competitive, achieving higher AUROC, F1, and balanced accuracy. These results highlight the importance of multi-revisit aggregation while motivating a closer examination of temporal modeling strategies.

Parameter-matched comparison. To assess whether performance differences arise from model capacity, we introduce a parameter-matched single-memory GRU (H259), containing 1,003,676 parameters compared with 1,004,175 for U$^2$-Revisit, a difference of less than 0.05%.

The two models achieve comparable Event AP (0.9596 vs. 0.9594), while U$^2$-Revisit obtains higher AUROC (0.9453 vs. 0.9440), F1 (0.9269 vs. 0.9175), and balanced accuracy (0.9013 vs. 0.8846). This comparison provides a capacity-controlled evaluation of the dual-memory design and its task-specific separation of spatial and identity representations.

Instance-memory ablation. Removing instance memory reduces mean Event AP from 0.9594 to 0.9549. We interpret this as a measurable but modest contribution to the joint learning framework, not as evidence that instance memory dominates event prediction.

These comparisons assess the complete task-adapted systems, with their respective supervision and implementation details documented in Appendix D.6.

Relevant results: Tables 3–4 and Appendices D.6–D.7.


AuthorsOpenReview ↗

Response to Reviewer JJQh (2)

CW3 & Q2: Statistical Significance and Cross-Site Variability

Concern: The dataset is small and heterogeneous, performance varies across sites, and the reported improvements may not be statistically reliable.

Response: We have added paired bootstrap confidence intervals to quantify uncertainty in the candidate-level Event AP comparisons.

The seven-site LOSO evaluation contains 544 labeled candidates from seven development sites, including 325 positives and 219 negatives. Reported model results are averaged over three training seeds.

Comparison (Event AP)Mean differenceCandidate-level 95% CISite-cluster 95% CI
U$^2$ vs. Unordered Set+0.0132[-0.0013, 0.0264][-0.0056, 0.0460]
U$^2$ vs. Matched GRU-0.0002[-0.0045, 0.0045][-0.0083, 0.0037]

The paired bootstrap analysis indicates that the Event AP differences between the strongest multi-revisit models are modest, with 95% confidence intervals spanning zero. The wider site-cluster intervals further reflect performance variability across evaluation sites.

To complement the seven-site LOSO evaluation and further assess generalization to unseen environments, we conducted additional experiments on two held-out sites, R10 and R11, comprising 211 labeled candidates:

MethodEvent APAUROC
Unordered Set0.9630.789
Single Memory0.9720.812
U$^2$-Revisit0.9730.817

The additional evaluation on R10 and R11 provides evidence of competitiveness on unseen sites, with the highest pooled Event AP (0.973) and AUROC (0.817) among the evaluated baselines. This evaluation contains 190 positives but only 21 negatives, so we treat it as additional competitiveness evidence rather than broad cross-domain robustness.

Relevant results: Site-wise analysis, Appendix D.5, and the paired bootstrap evaluation.


Q3: Controlled Missingness, Registration, Detector Failure, and Irregular Time Gaps

Question: Can robustness be evaluated under controlled missing observations, registration errors, detector failures, and changing revisit intervals?

Response: To evaluate robustness under incomplete observations, we introduced a controlled synthetic missing-evidence experiment at missingness rates of 10%, 30%, and 50%, using three independently trained model seeds and five perturbation seeds.

At selected revisit positions, visual and telemetry features and their availability information are masked, and the quality descriptors are set to zero. Static map features and original time intervals remain unchanged.

We compare two inference-time policies using identical perturbations:

  • Skip Update: Suppress the recurrent update at a masked position while retaining elapsed-time decay.
  • Forced Update: Perform the recurrent update despite the absence of dynamic evidence.
MissingnessSkip Update APForced Update APΔ\DeltaAP
10%0.95610.9522+0.0039
30%0.95370.9336+0.0200
50%0.94890.8926+0.0563

The advantage of Skip Update increases with the missingness rate, reaching +0.0563 Event AP at 50% missingness. This demonstrates the effectiveness of selective recurrent updates in preserving temporal memory under incomplete observations.

The complementary time-decay ablation further supports the proposed temporal modeling mechanism, with mean Event AP decreasing from 0.9594 to 0.9521 when time decay is removed.

We additionally evaluated 95 sequences containing naturally occurring detector-missing observations, identified from automatic high-recall records. U$^2$-Revisit achieves an Event AP of 0.7700, compared with 0.7563 for Unordered Set, providing complementary evidence under naturally occurring detection failures.

These experiments evaluate missing-evidence handling and elapsed-time modeling, while controlled registration perturbations and independently varied revisit intervals remain outside the current experimental scope.

Relevant results: Table 3 and Appendix D.8.


AuthorsOpenReview ↗

Response to Reviewer JJQh (3)

CW4 & Q4: Candidate-Proposal Recall and Selection Bias

Concern: The rule-based proposal stage may filter difficult events before classification, and evaluating only selected or labeled candidates could overestimate end-to-end performance. What is the proposal recall?

Response: We have clarified the distinction between candidate proposal and persistent-event verification in the revised manuscript. The proposal stage identifies region-centered sequences for review, while the primary evaluation measures the model's ability to verify persistent events using the available labeled evidence.

The seven-site benchmark comprises 650 region-centered sequences, of which 544 have determinate labels and are included in the candidate-level evaluation. The detailed composition is summarized below:

CategoryNumber
Total region-centered sequences650
Sequences selected by proposal stage362
Proposed sequences with determinate labels353
Additional evaluable negative sequences191
Sequences without determinate labels106
Total evaluable sequences544

Among the 353 proposed sequences with determinate labels, 325 are positive and 28 are negative. Including 191 additional negative sequences yields a total of 544 evaluable sequences, comprising 325 positives and 219 negatives.

Importantly, the proposal stage serves to identify candidate sequences for review, while final persistence labels are determined through manual assessment of multi-revisit evidence. This distinction separates candidate selection from event verification.

Regarding proposal recall, the current annotations do not exhaustively cover all unproposed sequences, so a reliable end-to-end recall estimate requires additional annotation. We have clarified this evaluation scope in the revised manuscript, with the reported Event AP measuring candidate-level persistent-event verification.

Manuscript Revision: Section 3.3 and Appendices A.4–A.5.


Closing Remarks

We thank the reviewer for the constructive feedback, which motivated a more comprehensive evaluation of U$^2$-Revisit.

The expanded experiments highlight the value of multi-revisit evidence and provide additional support for quality-aware fusion, elapsed-time modeling, and selective recurrent updates under missing observations. The published baseline comparisons, parameter-matched evaluation, and statistical analyses further strengthen the empirical assessment of our approach.

Together, these additions reinforce the contribution of region-centered persistent-event verification under intermittent aerial observations, supported by a dedicated benchmark and an evidence-aware temporal modeling framework.

We appreciate the reviewer's consideration of our revised manuscript and additional results.

We hope that our responses have addressed your concerns. If any questions or concerns remain, we would be happy to provide further clarifications and engage in further discussion. We sincerely appreciate your time and valuable feedback.

Reviewer_gi9jOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies persistent vehicle occupancy in fixed road and parking regions observed during intermittent UAV passes. It contributes an 11-site collection, with seven sites used for a site-disjoint leave-one-site-out (LOSO) benchmark, and distinguishes confirmed absence from unavailable or unreliable observations. A proposal stage selects candidate sequences, and U²-Revisit then combines visual, telemetry, map, and quality inputs using gated fusion and separate time-decayed spatial and instance memories. On 544 evaluable sequences, its pooled Event AP is 0.9594, compared with 0.9462 for an unordered set model and 0.8840 for a single-revisit model.

Critical strengths

1.The distinction between confirmed absence and unavailable evidence addresses a real failure mode of sparse aerial inspection. Fixed map regions also provide a sensible spatial unit when vehicle identity changes. 2.The evaluation uses held-out physical sites, discloses class counts for each site, reports three seeds and component ablations, and includes two additional external sites. These choices make the reported strengths and site variation easier to inspect

Critical weaknesses

1.The main decision does not use the vehicle memory. The model keeps one memory for the region and another for the vehicle. But the formula for deciding whether an event persisted reads only the region memory. As written, the vehicle memory cannot directly change that answer. It may still help indirectly during training. The paper should explain that role clearly, since the two-memory design is presented as a main part of the method 2.It is unclear whether a failed visit is counted when the model combines all visits for its final answer. If it is counted, adding a failed visit could change the answer even when the usable images stay the same. 3.The paper does not clearly say what counts as “persistent.” Suppose the UAV sees a car in the same place at 10:00 and 10:05. The car may have stayed there, or it may have left and returned between those visits. The paper does not give a clear rule for these cases, or say how many visits or how much time are needed for a positive label. Without that rule, readers cannot tell exactly what the model is being asked to predict.

Questions for authors

1.In Eq. 12, the final persistence score is calculated only from the region memory. Through which part of the model, if any, can the vehicle memory change that score? If its effect is only indirect, please clarify its role. 2.In Eq. 4, the fusion weights depend on quality and availability signals, rather than the visual, telemetry, or map features themselves. What are the five quality signals, and how should this gate handle two available inputs that give conflicting cues? 3.The paper treats “confirmed empty” and “detection missed” as different situations. What information does the model receive in each case, and how does it avoid treating a missed detection as an empty region? 4.In Eqs. 9–12, an unusable visit can leave the memory decayed and may still appear in the final summary. Are such visits excluded from that summary? If they are included, what should they contribute to the decision? 5.When a vehicle detection lies near the boundary between two fixed regions, how is it assigned to one region? 6.What exact rule makes a sequence a positive “persistent event”? Please clarify how many occupied visits are required, whether there is a minimum duration, and how gaps with no usable observations are treated. 7.A car seen in the same region on two visits may have left and returned while the UAV was away. How would this case be labeled? Does “persistent” mean occupancy at the observed visits, or uninterrupted occupancy between them?

3 comments
AuthorsOpenReview ↗

Response to Reviewer gi9j (1)

We thank the reviewer for the careful assessment and the specific questions about memory interaction, evidence quality, and the operational definition of persistent events. These questions are important for interpreting both the method and the benchmark. We address the three main concerns and seven questions below.

CW1 & Q1: Role of Instance Memory in Event Prediction

Concern: The instance-memory branch does not directly feed into the final persistent-event prediction head. How does it contribute to the primary task?

Response: We have clarified the distinct roles of the two memory branches in the revised manuscript. The candidate-level event prediction is derived from spatial memory, while instance memory supports identity-related auxiliary learning through shared feature representations.

The two memories have distinct roles:

  • Spatial memory maintains region-level occupancy evidence across revisits and provides the representation used for candidate-level event prediction.
  • Instance memory is trained through identity-related auxiliary supervision. Its contribution to the event task is indirect, through the jointly optimized shared feature representations, rather than a direct inference-time connection to the event head.

The ablation results show that removing the separate instance-memory branch reduces mean Event AP from 0.9594 to 0.9549, a difference of 0.0045. This is a modest effect, not evidence that the instance branch dominates event prediction.

To separate architectural design from parameter capacity, we additionally evaluated a parameter-matched single-memory GRU:

ModelParametersEvent APAUROCF1BAcc
Matched GRU (H259)1,003,6760.95960.94400.91750.8846
U$^2$-Revisit1,004,1750.95940.94530.92690.9013

The matched GRU achieves essentially identical Event AP, while U$^2$-Revisit performs better on the other three reported metrics. We therefore interpret the dual-memory structure as a task-motivated inductive bias with a modest measured contribution, rather than as an architecture with an established advantage over every capacity-matched alternative.

Relevant results: Table 3, Table 4, and Appendix D.7.


CW2 & Q4: Unavailable Revisits, Recurrent Updates, and Final Pooling

Concern: If a revisit is unusable, does its state still influence the final prediction through sequence pooling?

Response: This concern highlights an important distinction between recurrent state updates and final temporal aggregation. For a valid, non-padding revisit whose update mask is disabled, the model suppresses the recurrent update driven by unavailable dynamic evidence, while elapsed-time decay remains active. The resulting time-evolved memory state may still contribute to final temporal aggregation, preserving temporal information without incorporating unsupported dynamic evidence through a recurrent update.

The controlled missingness experiment directly evaluates this update policy. Both Skip Update and Forced Update use identical masked inputs and retain time decay.

MissingnessSkip Update APForced Update APΔ\DeltaAP
10%0.95610.9522+0.0039
30%0.95370.9336+0.0200
50%0.94890.8926+0.0563

The improvement increases with the missingness rate, reaching +0.0563 Event AP at 50% missingness.

Relevant results: Sections 5.2–5.4 and Appendix D.8.


AuthorsOpenReview ↗

Response to Reviewer gi9j (2)

Q2: Five Quality Signals and Conflicting Modalities

Question: What exactly are the five quality signals, and how are conflicting modality observations handled?

Response: The five quality descriptors summarize the reliability or availability of automatically extracted observations:

  1. Detection confidence: Confidence summary from vehicle detection.
  2. Visual availability: Whether usable visual features are available.
  3. Registration reliability: Quality of image-to-map alignment.
  4. GPS availability: Availability of the required coordinate fields.
  5. Blur quality: An image-sharpness descriptor based on Laplacian variance.

These descriptors, together with modality-availability indicators, inform the quality-aware fusion mechanism. The quality gate is not an explicit semantic contradiction-resolution module; it learns reliability-aware modality weights rather than applying hard conflict rules. When map context and visual detections suggest different interpretations, the model adjusts their relative contributions through learned reliability estimates instead of guaranteeing that a particular source prevails.

The quality-aware fusion mechanism adaptively weights different modalities according to their estimated reliability and availability, allowing the model to adjust their contributions when observational evidence is incomplete or inconsistent.

The component ablation supports the usefulness of this design: removing quality-aware fusion reduces mean Event AP from 0.9594 to 0.9482, a difference of 0.0112.

Relevant results: Section 5.3 and Table 3.


Q3: Confirmed Empty Versus Detection Missing

Question: How does the system distinguish a truly empty region from a region where the detector fails to provide usable evidence?

Response: We distinguish these two states based on the availability and reliability of observational evidence:

Confirmed empty: Usable observational support is available for an empty-state assessment.

Detection missing: Detection evidence is unavailable or insufficient, preventing a reliable assessment of whether the region is occupied.

The key distinction is that missing evidence does not imply an empty region. This allows the model to distinguish observed absence from insufficient observational evidence when reasoning across revisits.

Observation-state records, quality descriptors, and temporal update masks are generated through automatic preprocessing, without using manually annotated event or vehicle-identity labels as model inputs.

Manuscript Revision: Sections 3.3 and 5.2–5.3.


Q5: Vehicle-to-Region Assignment Near Boundaries

Question: How are detections assigned to regions when a vehicle lies close to a polygon boundary or overlaps multiple regions?

Response: We have audited the region-assignment procedure.

Each vehicle detection is associated with semantic regions using its projected bounding-box center in the reference-map coordinate system. The assignment is not based on maximum bounding-box overlap or polygon intersection area.

When the projected center falls inside multiple regions:

  1. A designated_parking region takes priority.
  2. Remaining ties are resolved by the original GeoJSON region order.

Each detection is assigned to at most one semantic region per frame. Region assignments are determined independently for each frame, allowing a tracklet to be associated with different regions over time.

This deterministic assignment strategy resolves overlapping polygon matches through predefined priorities, while localization and registration errors may still affect assignments near region boundaries.

Relevant discussion: Region association and preprocessing details in Appendix B.3.


AuthorsOpenReview ↗

Response to Reviewer gi9j (3)

CW3 & Q6: Operational Definition of Persistent Events

Concern: What qualifies as a positive persistent event? Is there a minimum number of occupied revisits or a minimum observed duration?

Response: We have clarified the operational definition of persistent events in the revised manuscript. A positive persistent event represents an occupancy episode of a fixed semantic region supported by evidence accumulated across multiple UAV revisits, rather than an isolated vehicle detection or continuous tracking of a particular vehicle.

Persistence is assessed through the available multi-revisit evidence. Repeated observations of regional occupancy support persistence, whereas confirmed-empty observations provide evidence of an interruption. Unavailable observations are treated as missing evidence rather than confirmed absence.

The proposal stage uses predefined rules to identify candidate sequences for manual review. The proposal criteria do not directly determine positive persistence labels, which are assigned based on the reviewed multi-revisit observations.

Regarding minimum revisit counts and durations, the current benchmark documentation does not establish a universal numerical threshold as a positive-label criterion. We therefore describe the evidence-based annotation procedure without introducing an unverified cutoff.

Manuscript Revision: Introduction, Section 3.3, and Appendices A.4–A.5.


Q7: Departure and Return During Unobserved Gaps

Question: If a vehicle leaves and another vehicle arrives before the next revisit, how is this case handled?

Response: This scenario illustrates the distinction between region-level persistence and vehicle-identity continuity in intermittent aerial observations.

When a region is occupied at two consecutive revisits, several transitions may have occurred during the unobserved interval: the same vehicle may have remained, departed and returned, or been replaced by another vehicle.

U$^2$-Revisit addresses these questions at two levels. The primary event prediction evaluates persistent occupancy evidence associated with the fixed semantic region, while the identity-related auxiliary tasks assess same-vehicle continuity and replacement using available visual and temporal evidence.

An intervening confirmed-empty observation provides evidence of an occupancy interruption, whereas an unavailable observation preserves uncertainty about the region's state. This separation allows the framework to reason about region occupancy and vehicle identity without treating them as equivalent.

Manuscript Revision: Section 3.3 and Appendices A.4–A.5.


Closing Remarks

We thank the reviewer for the constructive feedback, which has helped us clarify the roles of the two memory branches, the treatment of unavailable observations, and the operational definition of persistent events.

The revised manuscript provides clearer implementation details and additional experimental evidence supporting quality-aware fusion, temporal memory modeling, and selective recurrent updates under missing evidence. Together, these revisions strengthen the presentation and empirical foundation of U$^2$-Revisit.

We appreciate the reviewer's consideration of our revised manuscript and additional results.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their constructive feedback. The revised manuscript substantially strengthens the experimental evaluation and further clarifies our central contribution: region-centered persistent-event reasoning under intermittent aerial revisits, with explicit consideration of observation reliability and missing evidence.

Our work introduces a region-centered benchmark and U$^2$-Revisit, a quality-aware temporal reasoning framework designed for this setting. By integrating spatial correspondence across revisits, elapsed-time memory modeling, and evidence-aware feature fusion, we provide a systematic approach to verifying persistent events from sparse and potentially incomplete UAV observations.

We have added the following evidence:

Reviewer ConcernAdditional EvidenceManuscript Location
Limited published baselinesGRU-D, SeFT, ATENet, and MTMTable 4, Appendix D.6
Architectural contribution and capacityParameter-matched single-memory GRUAppendix D.7
Missing-evidence robustnessControlled missingness at 10%, 30%, and 50%Appendix D.8
Statistical uncertaintyPaired candidate- and site-level bootstrap analysesExperimental analysis
Cross-site generalizationAdditional evaluation on R10 and R11Appendix D.5
Task and annotation clarityClarified observation states, candidate accounting, and evaluation scopeSection 3.3, Appendices A.4–A.5

The expanded evaluation provides three key findings.

First, multi-revisit reasoning substantially improves persistent-event verification. U$^2$-Revisit achieves a mean Event AP of 0.9594, compared with 0.8840 for Single Revisit and 0.9462 for Unordered Set, demonstrating the value of aggregating evidence across intermittent observations.

Second, quality-aware temporal modeling contributes to robust event reasoning. Removing time decay reduces mean Event AP from 0.9594 to 0.9521. Under 50% controlled missingness, selective recurrent updates improve Event AP by 0.0563 compared with forced updates under identical perturbations. These results provide complementary evidence for temporal memory modeling and evidence-aware state updates.

Third, the expanded baseline evaluation positions the framework more precisely. U²-Revisit achieves higher Event AP than all four published task-adapted methods. The parameter-matched single-memory GRU obtains comparable Event AP, while U²-Revisit performs better on AUROC, F1, and balanced accuracy. We therefore present the dual-memory design as a task-motivated inductive bias and treat the benchmark, missing-evidence protocol, and quality-aware temporal formulation as the primary contributions.

Together, these additions strengthen the empirical foundation of region-centered persistent-event reasoning under intermittent aerial revisits. We have also clarified the candidate-level evaluation protocol and corresponding annotation procedures.

Key revisions are marked in blue in the updated manuscript for ease of reference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.