{"id":"ac1de3b2-cf79-4f34-8deb-981380bf2588","arxiv_id":"2506.18177","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A scalable belief-propagation track-before-detect method tracks multiple objects directly from correlated raw sensor data with fluctuating amplitudes and unknown noise, outperforming detect-then-track baselines.","lead":"This paper presents a Bayesian track-before-detect method that tracks multiple objects directly from raw radar or sonar measurements, without first converting them into detections. It uses a signal model with correlated measurements, fluctuating amplitudes, and unknown noise, and runs efficient belief propagation on a factor graph.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MP-seeded birth contradicts the detector-free TBD claim; the evaluated BP-TBD is a hybrid DTT/TBD method, so the central contribution is not demonstrated.","rationale":"The reader's weakest assumption identified the same issue: BP-TBD initializes new POs only on MP-detection cells, so weak targets without a detection cannot be born and the claimed TBD advantage is confined to detector-seeded regions. My analysis confirms this is the most load-bearing concern. It is not a minor implementation detail: it directly contradicts the abstract and introduction, and it changes the interpretation of every experimental result. The GOSPA improvements may be real, but they demonstrate a hybrid method that inherits track-initiation from a detector, not a detector-free TBD method. Other issues (missing error bars, Swerling-1 labeling, b_eta tuning) are secondary. The correct verdict is CONDITIONAL: the paper should not be accepted as supporting the stated central claim until full-grid birth is evaluated or the claim is revised to describe a detector-aided TBD method. I do not recommend REJECT because the underlying signal model and BP framework may still be valuable, and the full-grid variant could plausibly work; the evidence is simply incomplete.","tokens_in":22453,"tokens_out":2478,"duration_ms":31943,"concrete_test":"Re-run the synthetic radar experiment (Section IV-A) with the full birth model of Section II-C: initialize new POs on all Q=2500 grid cells (no MP seeding) while keeping all other parameters and the BP-TBD message-passing scheme unchanged. Report GOSPA over the same 100 runs at -6, -3, and 0 dB. If full-grid birth yields comparable or better GOSPA without MP detections, the detector-free claim is substantiated; if it fails computationally or underperforms because MP seeding is essential, then the central TBD claim is not supported by the paper as written.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that BP-TBD operates directly on raw sensor data 'without requiring a frontend object detector' and thereby preserves weak-object information that DTT discards. But the implemented method does exactly the opposite for track initiation. In Section IV-A (also IV-B and IV-C) the authors state: 'instead of initializing new POs for all grid cells, we only initialize new POs on cells with a MP detection. As a result, BP-TBD has an equal number of POs at each time step as MP + Tracking.' Since new potential objects are only born on cells where the matching-pursuit detector fires, any target that produces no MP detection is never represented by a PO and cannot be tracked. This is a DTT-style gating of track birth, not track-before-detect. The full-grid birth model of Section II-C is never evaluated, so the experiments provide no evidence that the method can detect targets missed by the frontend detector. The GOSPA comparison is therefore between two DTT variants (one with MP/SBL point measurements, one with raw-data updates but MP-seeded birth), not between TBD and DTT. The abstract's 'eliminating the need for an object detector' and the introduction's 'without requiring a frontend object detector' are not supported by the reported implementation. This is the load-bearing weakness: the distinctive claim of the paper—detector-free TBD—rests on a component that is, in practice, a conventional detector.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a track-before-detect (TBD) method for multi-object tracking based on a superpositional signal model that accounts for correlated sensor measurements, amplitude fluctuations (Swerling 1), and unknown noise power. The joint posterior is represented by a factor graph and solved with belief propagation, using particle representations and single-Gaussian moment matching for intractable messages. The method is evaluated on a synthetic multi-sensor radar scenario and two real passive-acoustic scenarios, comparing GOSPA against two detect-then-track baselines (MP+Tracking and SBL+Tracking). The paper claims that the proposed method 'eliminates the need for an object detector' and outperforms state-of-the-art conventional MOT methods.","tokens_in":22712,"tokens_out":5742,"duration_ms":61253,"significance":"If fully substantiated, the work would be a valuable contribution: it extends the superpositional signal model to multiple snapshots and dictionaries, explicitly handles correlated measurements without a conditional-independence assumption, and reports linear complexity in the number of potential objects. The real-data acoustic experiments are a useful addition to a literature dominated by synthetic studies. However, the central claim of detector-free TBD is not supported by the reported implementation, because track birth is seeded by a matching-pursuit detector rather than by the full-grid birth model. This issue is load-bearing and limits the significance of the experimental demonstration.","major_comments":[{"comment":"The implementation initializes new potential objects only on grid cells where the matching-pursuit detector fires, as stated: 'instead of initializing new POs for all grid cells, we only initialize new POs on cells with a MP detection.' This means the method still requires a frontend object detector for track initiation, contradicting the abstract's claim of 'eliminating the need for an object detector' and the introduction's 'without requiring a frontend object detector.' The full-grid birth model of Section II-C is described but never evaluated, so the experiments provide no evidence that the method can detect targets missed by the MP detector. The GOSPA comparison is therefore between two DTT-style variants (MP-seeded BP-TBD versus MP/SBL plus tracking) rather than TBD versus DTT. This is the central claim and it is not supported by the reported implementation.","section":"IV-A (also IV-B and IV-C)"},{"comment":"The covariance formulas used for the Gaussian-approximated messages are explicitly imported from previous work ('Based on derivations previously presented in [16], [22], [34]'), but no derivation or adaptation is provided for the new multi-snapshot, multi-dictionary, correlated-measurement model. Since these formulas are the core of the message computation, the paper should either (i) provide a derivation that shows how the correlated measurement model leads to these specific covariance expressions, or (ii) clearly delineate which aspects of the derivation are new. Without this, the novelty and correctness of the proposed inference method are difficult to assess.","section":"III-C, Eqs. (12)-(14)"}],"minor_comments":[{"comment":"The notation 'xk,n ∈ {1,...,Nk}' for the kinematic state is a typo; the state should be defined in the continuous state space, not as an index.","section":"II-A"},{"comment":"The expected number of new objects is set to µB = 10−6, which is effectively zero; the paper should clarify whether the full birth model is active at all in the synthetic experiments, given that MP-seeded initialization is used instead.","section":"IV-A"},{"comment":"The mechanism for initializing MP-seeded POs is not specified: it is unclear whether the method uses the birth model's prior fB(x, γ) for those cells, and with what existence probability. This should be clarified to make the implementation reproducible.","section":"IV-A (and IV-B, IV-C)"},{"comment":"The per-PO DFT bin selection based on estimated range rate is a potentially information-discarding preprocessing step; the paper should discuss how this interacts with the claim of operating directly on raw sensor data without a frontend detector.","section":"IV-C"},{"comment":"The paper does not report the number of initialized POs or computational runtimes, making it difficult to verify the scalability claim in practice.","section":"IV"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the mismatch between the detector-free claim and the MP-seeded birth implementation. This is fixable either by evaluating the full-grid birth model (at least in a synthetic scenario) or by substantially revising the claims to describe the method as a hybrid DTT/TBD approach. The authors also rely heavily on their own prior work for the covariance approximations; a more detailed derivation or clearer delineation of novelty would strengthen the paper. The manuscript is within the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the paper is worth reading, but don't trust the abstract's 'no frontend detector' claim. The implemented BP-TBD only creates new tracks on grid cells where a matching pursuit detector fires. That is a detector-gated birth, and since every experiment uses it, the reported results are for a hybrid DTT/TBD method, not a pure TBD one. The full-grid birth model is described in Section II-C but never evaluated. So the central claim that the method can detect weak objects that a frontend detector would miss is not demonstrated. This is the main weakness, and it is not minor for the paper's stated thesis.\n\nWhat is genuinely new: the measurement model. It drops the conditional independence across measurement elements that limits earlier BP-TBD (their own [16] and others), uses a full covariance with correlated entries, models amplitudes as zero-mean Gaussian with unknown power (Swerling-1 label aside), fuses multiple dictionaries and snapshots, and infers noise power online. That is a real step forward for superpositional TBD, and the factor-graph/BP derivation is careful and standard. The complexity claim, linear in the number of potential objects, is plausible. The real-data passive acoustic experiments are a plus; few TBD papers go beyond simulation.\n\nSoft spots worth naming:\n1. The birth issue above. The method should be evaluated with full-grid birth (or at least with a sensitivity analysis on birth seeding) before claiming detector-free operation.\n2. No comparison with any existing TBD method, including their own [16] or the Davies-García-Fernández method [22]. They say such methods are unsuitable for correlated measurements, but they could still run them on preprocessed data or compare on a simpler scenario. Without this, 'state-of-the-art' is only established against two DTT baselines.\n3. The Swerling 1 label looks wrong. Amplitudes are independent across snapshots, which is closer to Swerling 2. Fix the terminology or the model.\n4. The noise amplification factor bη is tuned per experiment (1, 10, 5). It's presented as a robustness knob, but it's still a free parameter. Report sensitivity.\n5. GOSPA curves have no error bars. Minor.\n\nNet: the modeling and inference machinery are solid, and the paper deserves refereeing. But the authors need to either tone down the 'no detector' claim or re-run with full birth, and they should benchmark against a TBD method. With that revision, this would be a useful contribution to the tracking community.\n\nRecommendation: send to peer review, but expect major revision. The central algorithm is worth engaging with even if the current framing overreaches.","headline":"Solid BP-TBD extension with a genuine modeling advance, but the 'detector-free' claim is undercut by MP-seeded track birth; needs a revised framing and TBD baselines before it can be called state-of-the-art.","tokens_in":23299,"tokens_out":3456,"would_cite":true,"duration_ms":39248,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A detector-free tracker beats detect-then-track on weak targets","keywords":["track-before-detect","multiobject tracking","belief propagation","factor graph","superpositional signal model","amplitude fluctuations","measurement correlations","passive sonar"],"falsifier":"Run the proposed BP-TBD on the same synthetic radar scenario but initialize potential objects on all grid cells (full-grid birth) instead of only on matching-pursuit detections, and compare the GOSPA; if the full-grid version fails to track weak objects or is computationally prohibitive, the claim that the method is detector-free TBD is weakened.","tokens_in":22157,"feed_emoji":"📡","tokens_out":4066,"duration_ms":42176,"temperature":0.7,"pith_summary":"This paper claims that track-before-detect (TBD) can work directly on raw, correlated sensor data without a frontend detector, by modeling the data as a superposition of object signals plus noise and doing Bayesian inference with belief propagation. The proposed model keeps correlations across measurement channels, lets signal amplitudes fluctuate over time, and treats noise power as unknown, so the likelihood reflects the physics of radar and sonar data more faithfully than earlier TBD models. Because the resulting factor graph is processed with scalable message passing, the method handles an unknown, time-varying number of objects. If correct, this removes two central obstacles to practical TBD: the need to assume independent measurements and the need to know signal amplitudes.","feed_headline":"A detector-free tracker beats detect-then-track on weak targets","feed_subtitle":"New BP-TBD method uses correlated raw measurements and fluctuating amplitudes; lower error on synthetic radar and real acoustic data.","key_machinery":"The mechanism is a factor graph representation of the joint posterior of the kinematic states, Bernoulli-Gaussian signal-power states, and noise-power states. The measurement likelihood factorizes over dictionaries and snapshots, with each factor a zero-mean complex Gaussian whose covariance contains the sum of object contribution outer products plus noise; this covariance structure carries the correlation information. Belief propagation on this graph replaces intractable joint marginalization with local message updates, and moment matching collapses the Gaussian-mixture messages to single Gaussians, so each update costs O($M^{3}$) and the overall cost is O($M^{3}$ I J N_k), linear in the number of potential objects.","core_discovery":"The paper's central claim is that a comprehensive superpositional measurement model accounting for correlated measurements, fluctuating complex amplitudes, unknown noise power, multiple snapshots, and multiple signal dictionaries can be paired with belief propagation to produce a scalable track-before-detect method that outperforms conventional detect-then-track approaches. The authors construct the joint posterior as a factor graph whose factors encode object dynamics, birth, signal-power states, and noise-power states, then compute marginal posteriors by iterated sum-product message passing. Intractable messages are approximated by particles or by single Gaussians obtained through moment matching, keeping the per-step cost linear in the number of potential objects, dictionaries, and snapshots. In experiments with synthetic LFMCW radar and real passive acoustic data, the method achieves lower GOSPA error than matching-pursuit and sparse-Bayesian-learning detectors paired with BP-based tracking.","pith_inferences":["The experimental reliance on matching-pursuit detections for birth seeding means the practical gain may be limited to refining tracks after a conventional detector has already found a seed; a full-grid birth variant would be the true test of detector-free TBD.","The same factor-graph and moment-matching machinery could extend to other superpositional sensing modalities, such as MIMO communications or distributed acoustic arrays, where correlation structure carries information.","The Bernoulli-Gaussian amplitude model could be extended to other fluctuation statistics or to amplitude correlations across snapshots if future applications require it."],"forward_implications":["TBD can operate directly on raw radar or sonar data with correlated channels, eliminating matched-filter preprocessing or decorrelation steps.","Unknown noise power is estimated online, making the method adaptable to changing noise levels and colored noise, as shown by the noise amplification factor bη > 1.","The multi-dictionary, multi-snapshot formulation makes the same method applicable to multi-sensor fusion, multi-frequency beamforming, and multi-waveform localization.","Linear scaling in the number of potential objects, dictionaries, and snapshots allows the approach to scale to large surveillance regions."],"supporting_citations":[{"why":"Supplies the amplitude-fluctuation likelihood model that the paper extends to multiple snapshots and dictionaries.","marker":"[32]"},{"why":"Prior BP track-before-detect method assuming conditional independence; the new model removes that assumption.","marker":"[16]"},{"why":"Provides the scalable BP multitarget tracking framework and the baseline MOT method used in comparisons.","marker":"[8]"},{"why":"Shows the BP-with-superpositional-model machinery on multipath SLAM, which the proposed method adapts to TBD.","marker":"[34]"},{"why":"Establishes the factor graph and sum-product algorithm used to represent the posterior and derive messages.","marker":"[44]"},{"why":"Particle belief propagation used to implement particle-based message representations.","marker":"[47]"},{"why":"Matching pursuit detector used both as a baseline frontend and to seed birth cells in the experiments.","marker":"[61]"},{"why":"GOSPA metric used to quantify tracking error in all experiments.","marker":"[62]"}],"fun_headline_variants":["Message passing TBD beats detect-then-track without a detector","Belief propagation track-before-detect handles sensor correlations","No detector frontend: message passing TBD uses raw measurements","Tracker that avoids detect-then-track wins on weak targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experimental demonstrations initialize new potential objects only on grid cells where a matching-pursuit detector fires, so the claimed detector-free advantage is not actually tested for object birth in regions where that detector produces no detection.","fun_headline_variants_meta":{"raw":{"variants":["Message passing TBD beats detect-then-track without a detector","Belief propagation track-before-detect handles sensor correlations","No detector frontend: message passing TBD uses raw measurements","Tracker that avoids detect-then-track wins on weak targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000496,"raw_usage":{"total_tokens":2462,"prompt_tokens":1005,"completion_tokens":1457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":1388}},"tokens_in":621,"tokens_out":1457,"duration_ms":13355,"temperature":1.0,"reasoning_tokens":1388,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:24:00.154541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed BP-TBD on the same synthetic radar scenario but initialize potential objects on all grid cells (full-grid birth) instead of only on matching-pursuit detections, and compare the GOSPA; if the full-grid version fails to track weak objects or is computationally prohibitive, the claim that the method is detector-free TBD is weakened.","supporting_citations":[{"cited_title":"Multitarget l ikelihood com- putation for track-before-detect applications with ampli tude fluctuations of type Swerling 0, 1, and 3,","cited_arxiv_id":null,"evidence_quote":"Supplies the amplitude-fluctuation likelihood model that the paper extends to multiple snapshots and dictionaries."},{"cited_title":"A BP method for t rack-before- detect,","cited_arxiv_id":null,"evidence_quote":"Prior BP track-before-detect method assuming conditional independence; the new model removes that assumption."},{"cited_title":"Message passing algorithms for scalable mult itarget tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the scalable BP multitarget tracking framework and the baseline MOT method used in comparisons."},{"cited_title":"Direct multipath -based SLAM,","cited_arxiv_id":null,"evidence_quote":"Shows the BP-with-superpositional-model machinery on multipath SLAM, which the proposed method adapts to TBD."},{"cited_title":"Fact or graphs and the sum-product algorithm,","cited_arxiv_id":null,"evidence_quote":"Establishes the factor graph and sum-product algorithm used to represent the posterior and derive messages."},{"cited_title":"Particle belief propagati on,","cited_arxiv_id":null,"evidence_quote":"Particle belief propagation used to implement particle-based message representations."},{"cited_title":"Matching pursuits with time-fr equency dic- tionaries,","cited_arxiv_id":null,"evidence_quote":"Matching pursuit detector used both as a baseline frontend and to seed birth cells in the experiments."},{"cited_title":"Gen- eralized optimal sub-pattern assignment metric,","cited_arxiv_id":null,"evidence_quote":"GOSPA metric used to quantify tracking error in all experiments."}],"review_version":1}