{"id":"ad8aedae-80d5-41b1-a147-9e1b4d26af12","arxiv_id":"2603.08457","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Entropy-driven adaptive camera–LiDAR selection in a particle filter yields a favorable accuracy–continuity trade-off for single-vessel coastal tracking versus fixed sensor policies.","lead":"A particle-filter tracker fuses shore camera and LiDAR and picks the more informative sensor each step using entropy reduction. Field tests on a marina boat suggest this adaptive choice balances near-range accuracy with longer-range continuity better than fixed single-sensor or always-on fusion.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Entropy-reduction proxy for modality selection is load-bearing yet unvalidated against actual tracking error in the reported field results.","rationale":"The reader correctly isolates the weakest link: the unvalidated assumption that entropy reduction computed from the particle filter’s predictive models is a faithful online proxy for real accuracy/continuity gains under maritime degradations. With only the abstract available (full manuscript body empty in the cacheable source), no metrics, likelihood equations, or ablation against oracle/heuristic selectors can be inspected, so the empirical systems claim remains unverifiable and the UNVERDICTED/LOW-confidence stance is appropriate. No stronger internal inconsistency or alternative load-bearing flaw is visible; the experimental framing (real testbed, GNSS GT, four configurations) is sound in principle. Therefore the stress-test does not alter the reader’s verdict or weakest-assumption diagnosis.","tokens_in":2394,"tokens_out":501,"duration_ms":9445,"concrete_test":"Extract per-bin predicted entropy reduction for each modality together with the realized position RMSE (vs GNSS) after the chosen update; compute Spearman/Pearson correlation and compare adaptive RMSE/continuity against an oracle that always picks the modality with lowest post-update error. If correlation < 0.4 or adaptive is statistically indistinguishable from a range-threshold heuristic, the IG proxy does not support the headline trade-off claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that the information-gain (entropy-reduction) adaptive policy yields a favorable accuracy–continuity trade-off versus fixed LiDAR-only, camera-only, and always-on fusion on a real marina track with GNSS ground truth. That claim requires the particle-filter predictive measurement models to produce an online entropy score that reliably ranks which modality will actually reduce posterior uncertainty under real degradations (LiDAR range dropouts, camera illumination/clutter). The abstract asserts switching “based on information gain” and gives only qualitative regime statements (LiDAR near-field, camera longer-range). No equations, particle likelihood forms, entropy computation details, correlation of predicted IG with realized RMSE/continuity, or oracle-selection baseline appear in the supplied text. If the modeled likelihoods systematically mis-match field returns, the adaptive policy could be no better than a simple range or availability heuristic, undermining the claimed practical baseline for resilient maritime surveillance.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes a particle-filter tracker for single-vessel tracking from fixed coastal platforms that performs sequential measurement-level camera–LiDAR fusion and selects, at each fusion time bin, the modality expected to yield the largest entropy reduction (information gain). It is evaluated in a real marina deployment (CMMI Smart Marina Testbed, Ayia Napa) with a shore-mounted 3D LiDAR, an elevated fixed camera, and onboard GNSS ground truth on a rigid inflatable boat. Four configurations are compared: LiDAR-only, camera-only, always-on fusion (“All sensors”), and the adaptive policy. The abstract reports that LiDAR dominates near-field accuracy, the camera sustains longer-range coverage when LiDAR returns become unavailable, and the adaptive policy achieves a favorable accuracy–continuity trade-off, positioning it as a practical sensor-selection baseline for resilient, resource-aware maritime surveillance.","tokens_in":2574,"tokens_out":970,"duration_ms":11841,"significance":"If the field results hold under transparent metrics and ablations, the work supplies a concrete, GNSS-grounded baseline for entropy-driven modality selection in coastal vessel tracking—an application where cameras and LiDAR fail in complementary regimes. The combination of sequential measurement-level fusion, an explicit information-gain selection rule, and a real marina deployment with external ground truth is practically useful for resource-aware maritime surveillance. The contribution is primarily systems-level rather than theoretical; its value rests on whether the adaptive policy is shown to outperform simple heuristics and always-on fusion with quantitative error, continuity, and selection statistics, not merely qualitative regime statements.","major_comments":[{"comment":"The central claim—that entropy-reduction selection yields a favorable accuracy–continuity trade-off—is load-bearing and currently under-supported by the available abstract. No RMSE/ATE (or equivalent) numbers, continuity/coverage fractions, trial counts, confidence intervals, or selection-frequency statistics are given for LiDAR-only, camera-only, All, and Adaptive. Without these, the “favorable trade-off” cannot be assessed or reproduced. The manuscript must report quantitative tracking error and continuity metrics for all four configurations, ideally with range-binned breakdowns that match the claimed near-field LiDAR / long-range camera regimes.","section":null},{"comment":"The information-gain score is the decision rule of the adaptive policy, yet the abstract provides no equations for the predictive measurement models, particle likelihoods, or the entropy (or expected entropy reduction) computation, nor any validation that predicted IG ranks modalities in the same order as realized error reduction under real degradations (range dropouts, illumination, clutter). A correlation of online IG with realized RMSE/continuity, or an oracle-selection baseline, is needed to show that the modeled likelihoods are a reliable proxy rather than an unvalidated heuristic. If the full text lacks this, it should be added; if present only qualitatively, it should be made quantitative.","section":null},{"comment":"Free parameters that materially affect both tracking and selection (process/measurement noise scales, particle count and resampling schedule, fusion/selection time-bin length) are not characterized in the abstract. Sensitivity of the adaptive vs. fixed-sensor ranking to these choices should be reported, or at least fixed and fully specified so that the comparison is reproducible. Without that, it is unclear whether the reported trade-off is robust or tuned to a particular operating point.","section":null}],"minor_comments":[{"comment":"Define “fusion time bin” and the decision interval for modality selection explicitly (duration, relation to sensor rates, and whether selection is exclusive or can include both).","section":null},{"comment":"Clarify what “All sensors” means operationally (simultaneous measurement-level update every bin vs. asynchronous fusion) so that the adaptive policy’s resource/accuracy comparison is well-posed.","section":null},{"comment":"State the vessel trajectory length, range envelope, number of runs/sessions, and environmental conditions (day/night, sea state, clutter) so that the qualitative near-field / long-range claims can be contextualized.","section":null},{"comment":"If figures or tables exist in the full manuscript, ensure they report error and continuity side-by-side for all four configurations and annotate modality switches against range or time.","section":null}],"recommendation":"major_revision","confidential_remarks":"Only the abstract was available in the review package (full manuscript body empty in the provided source). The recommendation of major_revision assumes the full paper exists and can supply the missing quantitative metrics, IG formulation, and validation of the entropy proxy; if the full text is similarly qualitative, the appropriate outcome would be reject or uncertain pending a complete submission. Scope is a reasonable fit for a robotics/maritime sensing venue if the empirical bar is met."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a field-validated single-vessel coastal tracker that fuses camera and LiDAR at the measurement level inside a particle filter and picks one modality per fusion bin by expected entropy reduction. The new piece is not the PF or IG machinery—those are standard—but the marina deployment (shore LiDAR + elevated camera, RIB with GNSS GT) and the four-way comparison (LiDAR-only, camera-only, always-on, adaptive) that reports a practical accuracy–continuity trade-off.\n\nWhat it does well: the problem framing is honest (illumination/clutter vs range dropouts), the testbed is real, and the qualitative regime story matches physics—LiDAR near, camera when LiDAR dies. Framing adaptive selection as a resource-aware baseline for maritime surveillance is the right product for this community. Circularity risk looks low: external GNSS and fixed baselines are the right checks.\n\nSoft spots, in proportion: the load-bearing assumption is that the PF’s predictive measurement models make entropy reduction a reliable proxy for which sensor will actually cut tracking error under real degradations. The abstract only asserts switching “based on information gain” and gives regime statements—no RMSE tables, continuity metrics, IG-vs-realized-error correlation, or oracle/heuristic ablations. Without those, adaptive could be little better than range or availability gating. Free parameters (noise scales, particle count, decision interval) will matter; they need to be reported and stress-tested. Citation pattern and math cannot be judged from what we have.\n\nWho it is for: people building shore-based multi-sensor maritime trackers who want a concrete selection baseline, not a theory paper. It deserves a serious referee if the full manuscript ships metrics, likelihood forms, and selection diagnostics. I would not put it in next week’s reading group unless someone is actively doing coastal fusion, and I would not cite it in my own work in the next year unless I am writing exactly in this niche. Send it to peer review; do not desk-reject on novelty alone—the deployment and comparison design earn referee time.","headline":"Real marina camera–LiDAR particle-filter demo with entropy-based modality switching; useful systems baseline, but the adaptive claim rests on an unshown IG–error link and we only have the abstract-level story here.","tokens_in":3233,"tokens_out":537,"would_cite":false,"duration_ms":10922,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"An entropy-driven policy that picks camera or LiDAR at each step in a particle filter delivers a better accuracy–continuity trade-off for single-vessel tracking than fixed single-sensor or always-on fusion.","keywords":["particle filter","sensor selection","camera-LiDAR fusion","entropy reduction","maritime tracking","adaptive sensing","information gain","vessel tracking"],"falsifier":"Re-run the same marina trajectory with identical particle-filter settings but force the adaptive policy to choose the modality that the information-gain score ranks second (or random); if accuracy–continuity then matches or beats the true adaptive policy, the entropy-proxy claim fails.","tokens_in":3263,"feed_emoji":"🛥️","tokens_out":836,"duration_ms":10136,"temperature":0.7,"pith_summary":"Coastal vessel tracking is hard because cameras fail under bad light and clutter while LiDAR fades with range and intermittent returns. This paper builds a particle-filter tracker that can fuse camera and LiDAR measurements sequentially and, more importantly, an adaptive sensing rule that at each fusion time chooses the modality expected to reduce uncertainty the most. In a real marina trial with shore-mounted sensors and GNSS ground truth on a rigid inflatable boat, LiDAR is strongest nearby, the camera keeps coverage farther out when LiDAR drops out, and the adaptive switcher sits between them with a favorable mix of accuracy and continuity. The result is offered as a practical, resource-aware baseline for resilient maritime surveillance from fixed platforms.","feed_headline":"Entropy picks camera or LiDAR for better vessel tracks","feed_subtitle":"Real marina trial: adaptive particle-filter fusion beats fixed single-sensor and always-on modes on accuracy and continuity.","key_machinery":"Information-gain (entropy-reduction) adaptive sensing policy inside a sequential measurement-level camera–LiDAR particle filter: at each fusion time bin it scores expected entropy reduction from each modality’s predictive measurement model and activates the most informative one.","core_discovery":"In a real shore-based marina deployment, an information-gain (entropy-reduction) adaptive sensing policy that selects the most informative modality—camera or LiDAR—at each fusion time bin inside a particle-filter tracker achieves a favorable accuracy–continuity trade-off relative to LiDAR-only, camera-only, and always-on multi-sensor fusion for single-vessel tracking.","pith_inferences":["The same entropy-selection rule could be applied to other complementary pairs (radar–camera, thermal–LiDAR) without redesigning the filter core.","If communication or power budgets are tight, the policy naturally yields a low-duty-cycle schedule that still preserves track continuity.","Multi-vessel scenes would stress the single-target particle representation and the scalar entropy score; an extension to multi-hypothesis or labeled filters is a natural next test.","Illumination or sea-state covariates could be folded into the predictive likelihoods to make the information-gain score more robust when models drift from field conditions."],"forward_implications":["Fixed coastal surveillance can keep near-field LiDAR precision while extending track life with the camera when LiDAR returns vanish.","Always-on dual-sensor fusion is not required for competitive performance; selective activation can save sensing and compute resources.","Entropy-based selection supplies a concrete baseline against which other maritime sensor-management policies can be compared.","The same particle-filter fusion stack can be re-used for other single-target coastal scenarios that share camera/LiDAR complementarity."],"fun_headline_variants":["Entropy selects camera or LiDAR for resilient vessel tracks","Adaptive modality pick yields better accuracy-continuity trade-off","Particle filter switches sensors via entropy for marina boat tracking","Info-gain policy blends camera and LiDAR for continuous vessel fix","Shore trial: entropy-driven fusion beats fixed single-sensor modes"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The entropy-reduction score computed from the particle filter’s predictive models is a reliable online proxy for which sensor will actually improve tracking under real maritime degradations.","fun_headline_variants_meta":{"raw":{"variants":["Entropy selects camera or LiDAR for resilient vessel tracks","Adaptive modality pick yields better accuracy-continuity trade-off","Particle filter switches sensors via entropy for marina boat tracking","Info-gain policy blends camera and LiDAR for continuous vessel fix","Shore trial: entropy-driven fusion beats fixed single-sensor modes"]},"model":"grok-4.5","effort":"low","cost_usd":0.005402,"raw_usage":{"total_tokens":1435,"prompt_tokens":754,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":54020000,"prompt_tokens_details":{"text_tokens":754,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":596,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":754,"tokens_out":85,"duration_ms":4724,"temperature":1.0,"reasoning_tokens":596,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T12:35:57.302732+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same marina trajectory with identical particle-filter settings but force the adaptive policy to choose the modality that the information-gain score ranks second (or random); if accuracy–continuity then matches or beats the true adaptive policy, the entropy-proxy claim fails.","supporting_citations":[],"review_version":1}