acceptodds
Under review as a conference paper at ICLR 2027

OmniGlimpse: Asymmetric Identity Memory for Online Multi-View Tracking

Abstract

Multi-view tracking extends visual coverage beyond a single camera, providing complementary observations when targets are occluded or leave the view. This raises a question: how can observations from different views be matched to the correct identities without offline scene-specific training or camera calibration? Progress toward this goal remains limited because many offline formulations assume such deployment-specific knowledge, while online adaptation can amplify association errors. We study this problem as online multi-view identity maintenance (OMIM): targets are initialized once per view, with no further identity labels provided after each initialization. To this end, we introduce OmniGlimpse, an asymmetric propose-verify framework that separates adaptive association from stable identity consolidation. At the representation level, short-term local and cross-view evidence is mapped into target-specific adaptive spaces to propose associations, whereas long-term identity distributions are maintained for verification in a canonical space defined by a frozen encoder. At the update level, short-term evidence may guide current associations, whereas, after bootstrapping, only verified observations may update persistent memory or drive adaptation. Under a unified target-specified evaluation protocol, OmniGlimpse improves over the best evaluated online baseline for each metric by 9.1 IDF1 and 22.1 CVIDF1 points on M3Track. The corresponding gains on MvMHAT are 20.3 and 26.2 points. Despite using no offline scene-specific training, it outperforms offline-trained references in IDF1 and remains competitive in cross-view association. These results support an approach to online identity maintenance that adapts to changing appearances while consolidating reliable evidence into stable identity anchors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi-view tracking, online identity maintenance, cross-view association, online adaptation, person re-identification

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

This paper proposes OmniGlimpse, an online multi-view tracking framework that separates adaptive association from stable identity verification. It employs target-specific adapters for online adaptation and long-term Gaussian identity distributions for verification. A verification-gated update mechanism is introduced to reduce identity drift. Experiments on M3Track and MvMHAT demonstrate improvements in identity maintenance and cross-view association.

Critical strengths
  1. The paper addresses an important challenge in online multi-view tracking by balancing adaptive representation learning with stable identity verification.
  2. The proposed framework achieves substantial improvements over the evaluated online baselines, with ablation studies supporting the effectiveness of its key components.
Critical weaknesses
  1. The experimental setting is relatively restricted, as each target is initialized in every camera view and only two designated targets are evaluated per scene. Moreover, the online baselines do not utilize cross-view information, which limits the fairness of the comparisons.
  2. The methodological novelty and necessity of the asymmetric dual-space design require further justification. Although the ablation studies demonstrate performance improvements, they do not fully establish its advantages over simpler alternative designs.
Questions for authors
  1. The evaluation assumes that each target is initialized in every camera view, with only two designated targets per scene. How well does the proposed framework generalize to scenarios with incomplete cross-view initialization or a larger number of targets?
  2. The online baselines operate independently across camera views, whereas OmniGlimpse utilizes cross-view information. Could the authors provide comparisons under more consistent cross-view information settings to better demonstrate the advantages of the proposed framework?
  3. The Adaptive Query ablation modifies only the verification query while retaining canonical Gaussian statistics. Could the authors further compare the proposed dual-space design with a unified adaptive feature space to better justify its necessity?
  4. Although the proposed method improves IDF1 and CVIDF1, Tables 1 and 2 show more identity switches than OCLReID. Could the authors further explain this observation and analyze its implications for long-term identity stability?
3 comments

We appreciate your positive assessment of the identity-memory design and your questions about its scope and supporting evidence. We have expanded the evaluation and clarified initialization and verification. Revisions are marked in blue.

Q1. Initialization setting and scalability to more targets. Related initialization protocols appear in prior work: Beyond SOT specifies multiple targets through initial bounding boxes [1], OCLReID adapts online to a designated person [2], and Multi-Drone-Based Single Object Tracking initializes the same target in each drone view before sharing target templates [3].

In OmniGlimpse, the initialization box at each target's first appearance in each view provides the only ground-truth identity supervision. Large viewpoint differences can make the same target look substantially different across cameras. Without local initialization, identity matching must rely on another view's appearance before a reliable local reference exists; an incorrect match can contaminate persistent memory and be reinforced by online updates. We therefore use one initialization per target per view as the minimal starting condition for our identity-memory and adaptation mechanism. No further identity annotation, scene-specific offline training, camera calibration, or geometric priors are required.

We also evaluated 2, 4, and 6 designated targets on the same eight M3Track scenes containing at least six valid targets. Comparing two with six targets, CVIDF1 remains comparable (54.52 vs. 54.60), as does MOTP (78.45 vs. 78.12). HOTA decreases from 58.36 to 52.53 and IDF1 from 77.11 to 69.65, but both change by less than 0.5 points from four to six targets. These results support tracking more designated identities while preserving comparable CVIDF1 and MOTP, with HOTA and IDF1 stabilizing at the higher target counts (Table 4, Section 4.4).

Q2. Cross-view baseline comparison. We added OCLReID+CV, which augments the original OCLReID pipeline [2] with target-specific cross-view memory, and the native multi-camera baseline OpenVINO MCMT [4]. The original view-independent online baselines also receive cross-camera target correspondence through aligned initialization labels, providing a favorable starting condition.

Cross-view memory improves OCLReID, but OmniGlimpse retains a clear performance advantage. Its IDF1 gains over OCLReID+CV are 6.03 points on M3Track and 17.38 points on MvMHAT, with corresponding CVIDF1 gains of 14.56 and 24.06 points. OpenVINO MCMT also performs substantially below OmniGlimpse on both datasets. Tables 1–2 contain the results, and Appendix B.2 describes the implementations.

Q3. Unified adaptive feature space. To address the potential query–memory mismatch in Adaptive Query, we added Unified Adaptive Space, using adaptive features for both verification queries and Gaussian memory. This variant improves CVIDF1 from 59.60 to 59.84 and reduces IDs from 44 to 41.

The full model further raises CVIDF1 to 61.29 and MDA from 53.58 to 55.03, while reducing IDs to 28. This comparison supports separating adaptive association from persistent verification: the adaptive representation accommodates current appearance changes, while the canonical representation keeps accumulated identity evidence in a fixed space. Table 3 and Section 4.3 report the comparison, with implementation details in Appendix B.3.

Q4. Identity switches and long-term identity stability. We expanded the analysis to examine target coverage, incorrect associations, error duration, and switch conditions (Tables 9–10, Appendix B.5). Relative to OCLReID, OmniGlimpse improves target recall from 61.27% to 70.83% on M3Track and from 37.75% to 61.93% on MvMHAT. It also substantially shortens incorrect-association segments, reducing their median duration from 8 to 3 frames and from 24 to 6 frames, respectively. These gains are accompanied by higher wrong-association rates (9.35% vs. 3.88% on M3Track and 9.38% vs. 7.00% on MvMHAT).

Occlusion or person overlap accounts for 57.1% of OmniGlimpse's switches on M3Track and 50.0% on MvMHAT, with most remaining switches associated with re-entry or small targets. After bootstrapping, observations that fail verification cannot update persistent Gaussian memory or supervise online adaptation, so a current association error does not automatically enter the persistent identity state.

References.

[1] Mayer et al. (2024). Beyond SOT: Tracking Multiple Generic Objects at Once. WACV.

[2] Ye et al. (2024). Person Re-Identification for Robot Person Following With Online Continual Learning. IEEE Robotics and Automation Letters.

[3] Zhu et al. (2020). Multi-Drone-Based Single Object Tracking With Agent Sharing Network. IEEE Transactions on Circuits and Systems for Video Technology.

[4] OpenVINO Toolkit. Open Model Zoo: Multi-Camera Multi-Target Tracking Python Demo.

Thank you for your response. While I consider this a solid piece of work, the novelty does not particularly impress me. Therefore, I have decided to maintain my current score while increasing my confidence rating.

Thank you for reading our response and considering the additional results. We appreciate your positive assessment and your comments, which have helped us further improve the paper. Thank you again for the time and attention you have given throughout the discussion.

Reviewer_YaX9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors argue that online model adaptation is vulnerable to erroneous tracking associations, which can introduce incorrect pseudo-labels and lead to identity drift. To address this issue, they propose an asymmetric propose–verify framework that separates adaptive target association from stable identity verification, enabling more reliable sample selection for online model updates.

Critical strengths

The authors raise a reasonable concern regarding the reliability of online model adaptation. Since erroneous tracking associations can introduce incorrect supervision, carefully selecting reliable training samples is important for maintaining identity consistency during online updates.

Critical weaknesses
  1. The proposed setting only requires tracking two pre-selected identities per scene, with ground-truth initialization in every view. I believe this substantially simplifies the identity association problem compared to general multi-view MOT. However, the authors do not sufficiently justify this restricted setting, and it remains unclear whether the proposed method can generalize to scenarios involving more target identities.

  2. A key weakness is the lack of strong baselines that are well matched to the proposed setting. In particular, all online baselines are run independently in each view without cross-view information exchange, whereas the proposed method explicitly exploits cross-view observations and online adaptation.

  3. In Table 3, the Adaptive Query variant shows a performance degradation compared to the full model. However, as explicitly acknowledged in Section 4.3 and Appendix B.3, this ablation only replaces the verification query with adaptive features, while the Gaussian memory remains in the canonical space. Therefore, the performance degradation may result from feature-space mismatch rather than demonstrating the necessity of the proposed asymmetric design. I believe a proper single-space baseline, where both queries and memory are maintained in the adaptive space, is necessary to validate this core design choice.

Questions for authors

The authors should address each concern raised in the Weaknesses section point by point. In particular, regarding the experimental setting, I strongly encourage the authors to provide relevant references demonstrating that similar settings have been adopted in prior research, rather than being an isolated and unconventional choice.

Additionally, I will carefully consider the perspectives of other reviewers, including both positive assessments and critical concerns, when determining my final evaluation of this paper.

5 comments

We appreciate your detailed assessment, particularly the concerns about the experimental setting, baseline comparability, and the verification-space ablation. We have added relevant references and experiments to address these points. Changes to the manuscript are highlighted in blue.

W1. Experimental setting and scalability to more targets. Target-specified tracking has precedents in prior work. Beyond SOT tracks multiple designated objects specified by bounding boxes in the initial frame [1], OCLReID maintains and re-identifies a designated person through online continual learning [2], and Multi-Drone-Based Single Object Tracking initializes the same target in the first frame of each drone view and shares target templates across views [3].

Following these settings, each designated target in OmniGlimpse is initialized once at its first appearance in each view. These initialization boxes provide the only ground-truth identity supervision, with no further identity annotations. The same target can appear substantially different across viewpoints, making initial identity matching based solely on another view's initialization unreliable. An incorrect initial match can contaminate persistent memory and be reinforced by subsequent online adaptation. We therefore consider one initialization per target per view the minimal starting condition for our per-view identity-memory and adaptation mechanism. OmniGlimpse requires no scene-specific offline training, camera calibration, or geometric scene priors.

To test scalability beyond two identities, we evaluated 2, 4, and 6 designated targets on the same eight M3Track scenes containing at least six valid targets. From two to six targets, CVIDF1 remains comparable (54.52 vs. 54.60), and MOTP is nearly unchanged (78.45 vs. 78.12). HOTA and IDF1 decrease from 58.36 to 52.53 and from 77.11 to 69.65, respectively, but change by less than 0.5 points each from four to six targets. These results demonstrate that OmniGlimpse supports more designated targets while maintaining comparable CVIDF1 and MOTP, with HOTA and IDF1 stabilizing at the higher target counts (Table 4, Section 4.4).

W2. Cross-view baseline comparison. Our protocol already benefits the original view-independent online baselines by providing cross-camera identity correspondence at initialization. To additionally control for access to cross-view observations during tracking, we evaluated OCLReID+CV, which extends the original OCLReID pipeline [2] with target-specific cross-view memory, and OpenVINO MCMT [4], a native multi-camera baseline.

OCLReID benefits from cross-view sharing, yet OmniGlimpse maintains a clear advantage over both added baselines. Compared with OCLReID+CV, its CVIDF1 is higher by 14.56 points on M3Track and 24.06 points on MvMHAT; MDA also improves from 38.31 to 55.03 and from 27.79 to 51.89, respectively. OpenVINO MCMT performs substantially below OmniGlimpse on both datasets. Thus, the gains persist when the comparison methods also have access to cross-view information. The results appear in Tables 1–2, with implementation details in Appendix B.2.

W3. Single-space adaptive verification. We agree that the original Adaptive Query control mixes adaptive queries with canonical Gaussian memory, so its degradation alone does not isolate the benefit of a fixed verification space. We therefore added Unified Adaptive Space, in which both verification queries and Gaussian identity memory use adaptive features.

This baseline improves over Adaptive Query, increasing CVIDF1 from 59.60 to 59.84 and reducing IDs from 44 to 41. The full model further improves CVIDF1 to 61.29 and MDA from 53.58 to 55.03, while reducing IDs to 28. These results support maintaining persistent identity verification in a fixed canonical space while using adaptive representations for association. The new comparison and analysis are included in Table 3 and Section 4.3, with implementation details in Appendix B.3.

References.

[1] Mayer et al. (2024). Beyond SOT: Tracking Multiple Generic Objects at Once. WACV.

[2] Ye et al. (2024). Person Re-Identification for Robot Person Following With Online Continual Learning. IEEE Robotics and Automation Letters.

[3] Zhu et al. (2020). Multi-Drone-Based Single Object Tracking With Agent Sharing Network. IEEE Transactions on Circuits and Systems for Video Technology.

[4]OpenVINO Toolkit. Open Model Zoo: Multi-Camera Multi-Target Tracking Python Demo.

I thank the authors for their response and the additional experiments. However, the proposed setting is considerably more restricted than general multi-view MOT: only a small set of pre-selected identities is tracked, and each target is initialized with a ground-truth bounding box in every view, thereby bypassing the initial cross-view identity discovery problem. Although the cited works establish precedents for individual aspects of this setting, none adopts the same inference protocol and evaluation strategy. I remain concerned about why this restricted formulation is necessary, rather than addressing a more general multi-view MOT setting, and whether the reported findings are representative of broader multi-view tracking scenarios.

Thank you for further clarifying this concern. Your question about the practical value of the setting and its connection to general multi-view MOT helps us explain the motivation and contribution more clearly.

Our setting targets applications where the identity of interest is known but must be maintained continuously. In video surveillance and security, person-search systems localize a specified target person from unconstrained scene images [5]. In embodied systems, long-term person following requires a mobile robot to track and re-identify a designated target over time [6]. In our setting, an operator provides target references, while the system must subsequently handle viewpoint changes, occlusion, re-entry, and appearance variation. OmniGlimpse requires no scene-specific offline training or camera calibration, reducing deployment-specific preparation.

This practical need motivates our core research question: why online adaptation driven by model-generated associations can still cause identity drift despite trusted initial references, and how such drift can be prevented. The references specify which targets to track, but subsequent detections must still be associated with the correct identities, and the system must decide which observations are reliable enough to update its representation. Incorrect associations can otherwise enter persistent memory and be reinforced through later online adaptation.

This problem is also relevant to broader multi-view MOT. Once cross-view identity correspondences are established, they must remain consistent over time, particularly when appearance or deployment conditions change and online adaptation is needed. For methods that update identity representations using their own predictions, preventing association errors from being amplified by learning is a shared challenge. Our target-specified formulation provides a defined identity reference for examining how association errors propagate through persistent memory and online adaptation.

Regarding target count, our additional experiments with 2, 4, and 6 designated targets demonstrate that OmniGlimpse extends beyond the original two-target setting (Table 4).

To further test whether OmniGlimpse depends on directly assigning a GT box to the target track in each view, we additionally evaluate shared reference-pool initialization. For each target, we construct a reference pool using the image crop from its first GT bounding box in each view. On M3Track, this gives two reference images per target, one from each view. Each view compares candidate detections against the shared reference pool and automatically determines detection-to-target and track-to-target assignments, without receiving GT box coordinates or GT-based track assignments for initialization. All subsequent tracking and online learning remain unchanged.

InitializationHOTA↑IDF1↑MOTA↑IDs↓MDA↑CVIDF1↑CVMA↑
Per-view GT boxes64.3883.3372.172855.0361.2966.77
Shared reference pool63.5482.2068.363145.4351.2955.92

With reference-pool initialization, HOTA and IDF1 remain close to the original protocol (63.54 vs. 64.38 and 82.20 vs. 83.33, respectively), with identity switches increasing from 28 to 31. Although this initialization introduces some degradation in cross-view metrics, the results demonstrate that OmniGlimpse also supports target specification through shared reference images, with detection matching and track assignment performed automatically. This extends its applicability beyond direct per-view box initialization and supports online identity maintenance under more flexible target-specification settings. Full details are provided in Appendix B.7.

References.

[5] W. Dong, Z. Zhang, C. Song, and T. Tan, “Bi-Directional Interaction Network for Person Search,” CVPR, pp. 2839–2848, 2020.

[6] X. Chen, J. Liu, J. Wu, C. Wang, and R. Song, “LoPF: An Online LiDAR-Only Person-Following Framework,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–13, 2022.

Reviewer_3XRUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online identity maintenance for designated targets across multiple camera views. Each target is initialized with a labeled bounding box at its first appearance in each view, with no further identity annotations thereafter. OmniGlimpse uses short-term local and cross-view evidence to propose associations, while long-term identity distributions in a frozen encoder’s feature space verify proposals and govern updates to persistent memory and target-specific adapters. Tables 1 and 2 report higher IDF1 and CVIDF1 than the evaluated online baselines. However, the initialization protocol and differences in access to cross-view information limit what can be concluded from these comparisons.

Critical strengths
  1. The paper clearly distinguishes evidence that is useful for a current association from evidence that is reliable enough to update the model. Verifying observations before persistent updates, while retaining a stable feature space, is a reasonable approach to limiting the propagation of erroneous online pseudo-labels.
  2. The reported IDF1 and CVIDF1 exceed those of the evaluated online baselines on both datasets. Table 3 also shows fewer identity switches when verification constraints are added to the online adaptation pipeline.
  3. The evaluation includes both within-view identity maintenance and cross-view association metrics. The paper also provides incremental ablations and controls that remove components from the full model.
Critical weaknesses
  1. Page 7 states that the online baselines process each view independently, whereas OmniGlimpse uses cross-view evidence. Consequently, the gains in Tables 1 and 2 do not, by themselves, establish the advantage of the asymmetric memory design. A stronger comparison would include an online baseline that can exchange cross-view features.
  2. The framework contains several components that may affect online efficiency, including target-specific adapters, short-term support banks, Gaussian identity prototypes, and verification-gated updates. The computational analysis is limited for a method intended for online tracking. Reporting the time spent on adaptation and classifier refitting would help establish its practical suitability. It would also be useful to show how these costs scale with the number of targets and camera views.
  3. The M3Track evaluation uses only RGB streams and excludes extremely low-light nighttime sequences, leaving 19 paired-view test instances. Although the reported comparisons follow a common evaluation protocol, this setting does not cover the full range of lighting conditions in the original benchmark. Since the method relies on appearance-based identification and online adaptation, evaluation under more challenging illumination would provide stronger evidence of its robustness and generalization.
Questions for authors
  1. The transition from “+Online Contrastive” to the full model in Table 3 introduces Gaussian verification together with associated update and learning constraints. While MDA and CVIDP improve, HOTA and IDF1 decrease. The paper should explain this trade-off and, if possible, isolate the effects of verification, update gating, and the distribution-anchoring loss.
  2. The identity-continuity results warrant further analysis. The full model incurs 28 identity switches on M3Track, compared with 18 for OCLReID, and 36 on MvMHAT, compared with 3 for OCLReID. Its CVIDF1 on MvMHAT is also lower than that of the offline reference CrossMOT. Please analyze where the switches occur, how long identity errors persist, and whether improved recall is accompanied by more incorrect associations, rather than relying mainly on aggregate IDF1.
2 comments

Thank you for the careful review and constructive suggestions on baseline fairness, online efficiency, and identity continuity. We have added cross-view comparisons, runtime profiling, low-light evaluation, ablations, and error analyses. Revisions are highlighted in blue.

W1. Cross-view baseline fairness. For the original view-independent online baselines, aligned initialization labels already supply cross-camera identity correspondence, giving them a favorable starting condition. To directly address access to cross-view information, we added OCLReID+CV, which retains the original OCLReID pipeline [1] while adding cross-view memory, and OpenVINO MCMT [2], a native multi-camera tracking baseline.

Cross-view sharing improves OCLReID on both datasets, but a clear performance gap remains to OmniGlimpse. Its CVIDF1 reaches 61.29 versus 46.73 on M3Track and 56.85 versus 32.79 on MvMHAT. OpenVINO MCMT also remains substantially below OmniGlimpse under the same protocol. These comparisons show that cross-view information alone does not account for the gains. Tables 1–2 report the complete results, with implementation details in Appendix B.2.

W2. Online efficiency and scalability. Table 5 reports total computation, adaptation, Ridge refitting, and memory usage across different target and view counts. Each view uses one NVIDIA A6000 GPU, with views processed in parallel. Adaptation and refitting costs are reported as per-frame averages.

Under two views, increasing the number of targets from two to six raises total computation from 189 to 369 ms/frame, adaptation from 32 to 86 ms/frame (approximately 17–23% of total computation), and GPU memory from 3.19 to 5.44 GiB. Ridge refitting remains lightweight at 3–10 ms/frame. At the same target count, moving from two to three parallel views adds 15 ms/frame in both tested configurations. Costs scale approximately linearly with target count, with limited additional overhead across parallel views.

W3. Low-light evaluation. We additionally evaluated the excluded extremely low-light M3Track sequences using RGB only and the same protocol. These sequences highlight the role of complementary infrared information when RGB appearance is severely degraded. Figure 4 shows representative frames at their original brightness.

Without scene-specific or nighttime training, OmniGlimpse improves over OCLReID in HOTA (22.18 vs. 18.34), IDF1 (30.72 vs. 24.00), MOTA (14.48 vs. 11.65), and CVIDF1 (14.83 vs. 10.98). The offline-trained CrossMOT reference remains stronger, indicating that extreme darkness is still challenging for the current RGB-only online setting. Table 8 and Appendix B.4 provide the complete results.

Q1. Verification, update gating, and distribution anchoring. We added two intermediate variants to separate these effects (Table 3; Appendix B.3). Starting from Online Contrastive, Gaussian verification reduces IDs from 83 to 61 and raises CVIDF1 from 58.50 to 59.61, with a small decrease in HOTA and IDF1. Verification-gated updates then reduce IDs to 39 and raise CVIDF1 to 60.72 by restricting persistent-memory admission and training eligibility. Finally, distribution anchoring reduces IDs to 28 and raises CVIDF1 to 61.29 while partially recovering HOTA and IDF1.

These results clarify the trade-off: online adaptation without Gaussian verification gives slightly higher conventional tracking accuracy, whereas verification and constrained updates improve cross-view identity consistency and reduce identity switches.

Q2. Identity continuity and incorrect associations. We added target recall, wrong-association rates, error duration, and switch-context analysis (Tables 9–10; Appendix B.5). Compared with OCLReID, OmniGlimpse increases target recall from 61.27% to 70.83% on M3Track and from 37.75% to 61.93% on MvMHAT. This higher recall is accompanied by higher wrong-association rates (9.35% vs. 3.88% and 9.38% vs. 7.00%, respectively), while median incorrect-association durations decrease from 8 to 3 frames and from 24 to 6 frames.

Occlusion or person overlap accounts for 57.1% of OmniGlimpse's switches on M3Track and 50.0% on MvMHAT. Most remaining switches occur during re-entry or when targets are very small.

On MvMHAT, the offline-trained CrossMOT reference achieves higher CVIDF1 (60.78 vs. 56.85) and fewer identity switches (17 vs. 36). OmniGlimpse achieves higher target recall (61.93% vs. 58.03%) together with a lower wrong-association rate (9.38% vs. 10.45%), showing that its higher coverage is not accompanied by a higher wrong-association rate relative to CrossMOT.

References.

[1] Ye et al. (2024). Person Re-Identification for Robot Person Following With Online Continual Learning. IEEE Robotics and Automation Letters.

[2] OpenVINO Toolkit. Open Model Zoo: Multi-Camera Multi-Target Tracking Python Demo.

Thank you again for your detailed review. In our initial response, we provided additional experiments and analyses to address your concerns. These additions have been incorporated into the revised manuscript.

Since then, we have also added an experiment with shared reference-pool initialization. For each target, we construct a shared reference pool using one target image from each view. Each view then matches detections against this pool and automatically assigns tracks to the target, without direct per-view GT track assignment. These results show that OmniGlimpse supports more flexible target specification through shared reference images, extending beyond direct per-view box initialization. The setup and results are provided in Appendix B.7 (Table 11).

We would appreciate any feedback on the revised results or our earlier response.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses

We thank the Area Chair and reviewers for their careful evaluation. We have revised the manuscript to clarify the experimental setting, strengthen the comparisons, and further examine the main design choices. Revisions are highlighted in blue.

We clarified the target-specified online setting and its relation to prior work. Each target is initialized once per view, with no further identity supervision. We additionally evaluated 2, 4, and 6 designated targets on the same M3Track scenes and profiled runtime across different numbers of targets and views (Tables 4–5).

We added OCLReID+CV, which extends OCLReID with cross-view memory, and the native multi-camera baseline OpenVINO MCMT on both datasets (Tables 1–2). OmniGlimpse maintains a clear advantage over both baselines, with CVIDF1 gains over OCLReID+CV of 14.56 points on M3Track and 24.06 points on MvMHAT. These comparisons show that access to cross-view information alone does not account for its gains. We also evaluated the extremely low-light M3Track sequences under the same RGB-only protocol and added representative nighttime examples (Table 8 and Appendix B.4).

The expanded ablation study separately evaluates Gaussian verification, verification-gated updates, and distribution anchoring. A new Unified Adaptive Space baseline uses adaptive features for both verification queries and Gaussian memory; the full model achieves higher CVIDF1 and fewer identity switches (Table 3). Finally, the added identity-continuity analysis examines target recall, incorrect associations, error duration, and switch conditions, clarifying the trade-offs between target coverage and identity stability (Appendix B.5).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.