{"id":"606b5465-3c11-420e-8bf4-44753061a3ba","arxiv_id":"1908.03945","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A HISP filter tracker with deep appearance features (HISP-DAL) reaches 37.4 MOTA on MOT16 and 45.4 MOTA on MOT17 using public detections.","lead":"This paper builds an online multi-person tracker from the HISP filter, a Bayesian filter that preserves identities with linear complexity, and adds deep-learning appearance features. It reports improved accuracy on the MOT16 and MOT17 benchmarks over its motion-only predecessor, though the gains are modest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The deep-appearance likelihood in Eqs. (11)–(13) has no defined track appearance state or template update, so the reported MOTA gain cannot be traced to the equations as written.","rationale":"The reader's weakest assumption is the same one that this pass finds load-bearing: the augmented likelihood cannot be evaluated without an appearance state/template that the paper never defines. I agree with that identification and add one detail: Eq. (13) is not only underspecified but internally inconsistent as printed, since the integrated likelihood retains Hx on the right-hand side; this strengthens the need for code or a formal correction. This does not change the direction of the reader's verdict: the condition is genuine and addressable, so the appropriate verdict remains conditional rather than accept or reject. If the requested instrumented ablation is provided and the appearance store is documented, the concern dissolves and the claim could be accepted; if no such store exists, the verdict should move to unverified or reject. The paper has independent support in the form of public benchmark evaluation and a defined network training procedure, but those do not repair the missing state definition in the filter equations.","tokens_in":22114,"tokens_out":6530,"duration_ms":72602,"concrete_test":"Ask the authors for the implementation, or reimplement from the text, and instrument the HISP update to log, for every hypothesis and frame, the exact source of z_iw in Eq. (21): which stored template or appearance state provides it, when it was created, and how it is updated after an assignment or a miss-detection. Then rerun the MOT16 DPM ablation twice with all else equal: once with Eq. (21) active and once with g_t(z_w|x)=1. If the code contains no appearance state or template store, or if the 1.5-point MOTA gap (35.9 vs 37.4) does not reproduce, the central claim is unsupported; if the gap reproduces with a documented store, the paper only needs to add that mechanism to Section III/V.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adding deep appearance information through Eqs. (11)–(13) improves tracking from 35.9 (HISP-T) to 37.4 MOTA (HISP-DAL). For this to be true, g_t(z_w|x) must be evaluable for every hypothesis during the update. The paper does not define the object needed to evaluate it. Section V defines the Kalman state as x_t=[p_cx, p_cy, v_x, v_y, w, h]^T and the observation as [p_cx, p_cy, w, h]^T; no appearance descriptor is in the state, and no initialization, propagation, or update of a track appearance template is given. Yet Eq. (21) compares features z_iw and z_jw 'extracted from the detected object regions determined by the detection vectors z_i_d and z_j_d', and Eq. (13) says z_i_d is the predicted measurement Hx. Extracting a feature from a predicted Hx box requires cropping the current frame at a predicted location and feeding it to VerIdNet; after a miss-detection or on the first frame of a track this is not well-defined without an external store, and no update rule for such a store appears anywhere. Eq. (13) also retains Hx on the right-hand side after the integral over x, which for a Gaussian p(x) should be the predicted mean Hm, suggesting the likelihood was never fully worked out. At minimum, the ablation cannot separate the effect of the deep features from the unstated bookkeeping actually used in the implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online multi-target visual tracker based on the Hypothesized and Independent Stochastic Population (HISP) filter, combined with a deep CNN appearance model (VerIdNet). The authors augment the HISP measurement likelihood with an appearance term, g_t(z|x)=g_t(z_d|x)g_t(z_w|x), and evaluate the resulting tracker (HISP-DAL) on MOT16 and MOT17. They report MOTA of 37.4 on MOT16 (versus 35.9 for the motion-only HISP-T) and 45.4 on MOT17, and claim the deep appearance information is responsible for the improvement. The paper also introduces a heuristic to re-label spawned tracks that share an identity after track extraction.","tokens_in":22536,"tokens_out":3245,"duration_ms":36308,"significance":"If the central claim is correct, this is the first HISP filter with deep representation learning and a working online identity-preserving tracker, which would be a useful contribution to visual multi-target tracking: the HISP filter offers linear complexity while maintaining track identities, and the addition of discriminative appearance features is a natural and potentially valuable extension. The benchmark results are self-consistent, and the use of external MOT16/MOT17 evaluation reduces circularity concerns. However, the current manuscript does not fully specify the mechanism by which the deep appearance likelihood is evaluated, and one key equation is mathematically problematic, so the central claim is not yet verifiable from the text.","major_comments":[{"comment":"The augmented likelihood g_t(z_w|x) in Eq. (21) cannot be evaluated as written because the paper never defines a track appearance state. Section V defines the Kalman state as x_t=[p_cx, p_cy, v_x, v_y, w, h]^T and the measurement as [p_cx, p_cy, w, h]^T, with no appearance descriptor. Eq. (21) requires z_i_w, an appearance feature extracted from the detected object region determined by the predicted measurement Hx, but no template store, no initialization rule, and no update rule for an appearance template are given. Please specify how the appearance feature of a predicted track is obtained: is a crop taken from the current frame at the predicted box, or is a stored template used? How is the template initialized at track birth and updated after association, and how are miss-detections handled? Without this, the reported MOTA gain from deep appearance cannot be traced to the equations provided.","section":"§III-B and §IV-A"},{"comment":"Equation (13) is not a valid identity. The left-hand side is an integral over x, but the right-hand side retains Hx and g_t(z_w|x) after the integral. For a Gaussian p(x), the detection-term integral should be evaluated at the predicted mean Hm, and the appearance term, if it depends on x, must remain inside the integral or be approximated explicitly. As written, the equation cannot be used to compute the association weight w^{κ,z}_t. Please correct the derivation and state the approximations used to factor the appearance term out of the integral.","section":"§III-B, Eq. (13)"},{"comment":"The ablation compares HISP-DAL with HISP-T, which is reported in the author's prior work [22], rather than a re-run of the same implementation without the appearance term. The paper states that HISP-T is 'simply the HISP-DAL tracker without deep appearance information,' but no evidence is provided that all other components (Kalman parameters, pruning/merging thresholds, track extraction window, detection preprocessing) are identical. The 1.5 MOTA difference could stem from implementation differences or parameter changes. Please provide a controlled ablation: run the same code with the appearance likelihood disabled, or report the exact configuration differences between HISP-DAL and HISP-T.","section":"§VI, Ablation Study"}],"minor_comments":[{"comment":"The abstract claims the tracker 'significantly outperforms several state-of-the-art trackers in terms of tracking accuracy,' but on MOT16, AMIR and KCF16 achieve higher MOTA (47.2 and 48.8 versus 37.4, Table I). The claim is acceptable if 'several' is interpreted as 'many, but not all,' but it would be clearer to qualify the statement or highlight the MOT17 results where the tracker does rank first in MOTA.","section":"Abstract and §VI"},{"comment":"The detection likelihood in Eq. (12) is written as an unnormalized exponential. Since the text states that the augmented likelihood is 'dimensionless and takes values in the interval [0,1]', the normalization is not just cosmetic; please define whether g_t(z_d|x) is a likelihood ratio, a density, or an unnormalized score, and clarify how it is normalized.","section":"§III-B, Eq. (12)"},{"comment":"Eq. (22) defines C_w as a cosine similarity (dot product divided by norms), but the text calls it a cosine distance. A similarity of +1 corresponds to identical features, while a distance would be 0 in that case. Please make the terminology consistent and state the range of C_w, since Eq. (21) is a sigmoid of 2C_w and the interpretation depends on the sign convention.","section":"§IV-A, Eqs. (21) and (22)"},{"comment":"The proposed re-labeling of spawned tracks based on the maximum-weight assumption is described as rarely susceptible to identity switches, but no quantitative evidence supports this claim. A brief analysis of how often the assumption fails (e.g., on the reported sequences) would help the reader assess the robustness of this post-processing step.","section":"§III-D"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the underspecified appearance mechanism. If the author can provide a precise specification of the appearance state (or point to code), and correct Eq. (13), the paper would be much stronger. The comparison tables are mostly self-consistent, but the ablation against a prior paper rather than a controlled re-run is a significant weakness. I would not recommend rejection if the specification can be supplied, but it is essential for the central claim to hold."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of the HISP-DAL paper. The core result is a real integration of the HISP filter with a deep appearance model, and the ablation against the author's own motion-only HISP-T is the most useful bit: on MOT16, MOTA goes from 35.9 to 37.4, and IDF1 from 28.9 to 30.5. That's modest but in the right direction, and it is measured on a public benchmark, not self-made data. The VerIdNet training on a large re-ID corpus is also sensible, and the paper does a decent job of explaining the HISP machinery for a vision audience. \n\nThe soft spots are real but not fatal. The big one is that the augmented likelihood is not fully specified. Equation (13) writes an integral over x but ends up with g_t(z_w|x) sitting outside the integral, with x still on the right-hand side. The text says z_i_d is the predicted measurement Hx, and the state vector is kinematic plus box dimensions; there is no appearance descriptor in the state, no initialization or update rule for an appearance template. I can see a plausible implementation: crop at the predicted box location, extract a VerIdNet feature, compare with the detection candidate. But that has to be stated. A referee will rightly ask how the crop is handled after a miss-detection or on the first frame, and whether the appearance comparison actually uses the predicted box or something else. \n\nThe abstract says 'significantly outperforms' several SOTA trackers. On MOT16, AMIR and KCF16 both beat HISP-DAL on MOTA, so that phrasing is too strong. The 3.3 fps runtime also doesn't support 'real-time' claims. And with no code or error bars, the single-run ablation is hard to trust fully. \n\nNone of this sinks the main claim. The integration is new, the benchmarks are standard, and the measured improvement is internally consistent. The paper deserves a serious referee; it needs a revised likelihood section and a code release to be fully convincing. I'd send it to review with a request for minor-to-major revision.","headline":"A plausible HISP-plus-deep-appearance integration with modest but consistent MOT gains, undercut by an underspecified likelihood and no code.","tokens_in":23010,"tokens_out":3611,"would_cite":true,"duration_ms":38337,"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 HISP filter with a deep appearance likelihood lifts MOTA from 35.9 to 37.4 on MOT16 and to 45.4 on MOT17 while keeping track identities online.","keywords":["HISP filter","stochastic populations","multi-target tracking","deep appearance learning","tracking-by-detection","person re-identification","MOT benchmark","augmented likelihood"],"falsifier":"Rerun HISP-DAL on the MOT16 test set with the appearance factor $g_t(z_w\\mid x)$ in Eq. (21) replaced by the constant 1: if MOTA remains 37.4, the deep appearance term is not carrying the reported improvement.","tokens_in":21937,"feed_emoji":"🎯","tokens_out":10955,"duration_ms":103770,"temperature":0.7,"pith_summary":"This paper tries to establish that the HISP filter, a stochastic-population multi-target filter that carries track identities at linear computational cost, can serve as an online visual tracker when its measurement likelihood is augmented with deep CNN appearance features. The author trains a Siamese ResNet50 verification-identification network (VerIdNet) on 6,654 person identities and combines the cosine distance of its features with a Gaussian motion likelihood inside the filter's hypothesis weights. The central reported evidence is that the deep-appearance version (HISP-DAL) raises MOTA from 35.9 to 37.4 and IDF1 from 28.9 to 30.5 on MOT16 relative to the motion-only HISP-T, and reaches 45.4 MOTA on MOT17 with public detections. If correct, this is the first HISP-filter tracker with deep representation learning, and it shows that appearance cues can enter stochastic-population filters through the likelihood rather than through a separate association stage.","feed_headline":"Deep appearance lifts HISP tracker to 45.4 MOTA on MOT17","feed_subtitle":"CNN features in the identity-preserving HISP filter beat its motion-only version and rival offline trackers on MOT benchmarks.","key_machinery":"The load-bearing object is the HISP filter itself, a Bayesian multi-target filter that represents each previously detected target by a measurement path, or hypothesis, and propagates each hypothesis' probability density and survival weight with complexity linear in the number of hypotheses and measurements. Into this filter the paper inserts an augmented likelihood of the form $g_t(z|x)=g_t(z_d|x)g_t(z_w|x)$: the detection term is the Gaussian innovation of Eq. (12), and the appearance term is the sigmoid of the cosine distance between CNN features, Eq. (21). The VerIdNet supplies those features: a Siamese ResNet50 with one verification and two identification losses, trained on 6,654 identities. The filter's update equations then convert these likelihoods into association weights and posterior hypothesis weights, so the appearance term is used both for data association and for track updates.","core_discovery":"The paper's central assertion is that the HISP filter's per-track hypotheses can ingest an augmented likelihood $g_t(z|x)=g_t(z_d|x)g_t(z_w|x)$, where the first factor scores box-level agreement through the Kalman innovation and the second factor scores appearance agreement through a logistic function of the cosine distance between deep CNN embeddings. With this likelihood, HISP-DAL reports 37.4 MOTA and 2,101 identity switches on MOT16 versus 35.9 MOTA and 2,592 switches for the motion-only HISP-T, and reaches 45.4 MOTA on MOT17. The paper also claims that the weight propagated with each confirmed hypothesis can relabel spawned objects that would otherwise share one identity after track extraction. The appearance embeddings come from VerIdNet, a Siamese ResNet50 trained with two identification losses and one verification loss on a mixture of re-identification and tracking data sets, which the paper reports at about 98% pair-verification accuracy.","pith_inferences":["The paper leaves a track's appearance state undefined: nothing in the Kalman state of Section V stores or updates an appearance template, so a working implementation must add an undocumented appearance store. Making that store explicit would also make Eq. (13) testable.","The same cosine-sigmoid likelihood could be ported to other Bayesian multi-target filters, such as labeled random-finite-set filters; if the reported gain is real, those filters should see a similar improvement.","The VerIdNet's 98% pair-verification accuracy contrasts with the tracker's 30.5 IDF1 on MOT16, suggesting that the appearance signal is diluted by the motion term or by association ambiguity; a testable extension is to reweight $g_t(z_d|x)$ and $g_t(z_w|x)$.","Since HISP-DAL is compared only with the author's own motion-only HISP-T, an independent appearance-based tracker under identical detections would be needed to isolate the contribution of the deep features."],"forward_implications":["An online tracker can preserve track identities without exponential hypothesis growth; the reported complexity is $O(|\\mathcal{I}_{t|t-1}||Z_t|)$ per recursion.","Appearance learning can be added to stochastic-population filters through the measurement likelihood, so the same augmented likelihood should extend to DISP and related filters, not only HISP.","The paper's reported MOTA on MOT17 is obtained with public detections from DPM, FRCNN, and SDP detectors, so the gain is not tied to a custom detector.","The duplicate-label fix is applied after track extraction and does not alter the filtering recursion, which means any HISP implementation can adopt it without changing the filter itself."],"supporting_citations":[{"why":"Originates the HISP filter and its linear-complexity recursion, which this paper applies to video.","marker":"[14]"},{"why":"Presents the author's motion-only HISP tracker (HISP-T), the ablation baseline for the deep-appearance version.","marker":"[22]"},{"why":"The author's GM-PHD filter with deep appearance learning, whose likelihood design and MOT results this work extends and compares against.","marker":"[3]"},{"why":"Provides the verification-identification CNN embedding and pair-sampling schedule used to train VerIdNet, and the 86% accuracy comparison point.","marker":"[20]"},{"why":"Supplies the ResNet50 architecture used as the backbone of VerIdNet.","marker":"[18]"},{"why":"Supplies Market1501, the largest person re-identification dataset in the training mixture, and the training set for the comparison model.","marker":"[39]"},{"why":"Defines the MOT16/MOT17 benchmark, its public detections, and the CLEAR and identity metrics used in the evaluation.","marker":"[47]"},{"why":"MHT-DAM is the appearance-based offline tracker that provides a strong reference comparison in both result tables.","marker":"[10]"}],"fun_headline_variants":["HISP+deep CNN hits 45.4 MOTA on MOT17","Deep appearance boosts HISP tracker to 45.4 MOTA","Identity-preserving filter with CNN features tops MOT17","VerIdNet embeddings lift HISP filter to 45.4 MOTA","HISP-DAL: deep likelihood beats motion-only on MOT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The augmented likelihood can be evaluated only if every track has an appearance feature vector available at its predicted state, but the paper never defines how such a track appearance template is initialized, stored, or updated.","fun_headline_variants_meta":{"raw":{"variants":["HISP+deep CNN hits 45.4 MOTA on MOT17","Deep appearance boosts HISP tracker to 45.4 MOTA","Identity-preserving filter with CNN features tops MOT17","VerIdNet embeddings lift HISP filter to 45.4 MOTA","HISP-DAL: deep likelihood beats motion-only on MOT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000612,"raw_usage":{"total_tokens":2836,"prompt_tokens":921,"completion_tokens":1915,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":1824}},"tokens_in":537,"tokens_out":1915,"duration_ms":13604,"temperature":1.0,"reasoning_tokens":1824,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:57:45.294105+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun HISP-DAL on the MOT16 test set with the appearance factor $g_t(z_w\\mid x)$ in Eq. (21) replaced by the constant 1: if MOTA remains 37.4, the deep appearance term is not carrying the reported improvement.","supporting_citations":[{"cited_title":"Multitarget filtering with linearized complexity,","cited_arxiv_id":null,"evidence_quote":"Originates the HISP filter and its linear-complexity recursion, which this paper applies to video."},{"cited_title":"Online multi-target visual tracking using a HISP filter,","cited_arxiv_id":null,"evidence_quote":"Presents the author's motion-only HISP tracker (HISP-T), the ablation baseline for the deep-appearance version."},{"cited_title":"Online multi-object visual tracking using a GM-PHD filter with deep appearance learning,","cited_arxiv_id":null,"evidence_quote":"The author's GM-PHD filter with deep appearance learning, whose likelihood design and MOT results this work extends and compares against."},{"cited_title":"A Discriminatively Learned CNN Embedding for Person Re-identification","cited_arxiv_id":"1611.05666","evidence_quote":"Provides the verification-identification CNN embedding and pair-sampling schedule used to train VerIdNet, and the 86% accuracy comparison point."},{"cited_title":"Scalable person re-identification: A benchmark,","cited_arxiv_id":null,"evidence_quote":"Supplies Market1501, the largest person re-identification dataset in the training mixture, and the training set for the comparison model."},{"cited_title":"Multiple hypothesis tracking revisited,","cited_arxiv_id":null,"evidence_quote":"MHT-DAM is the appearance-based offline tracker that provides a strong reference comparison in both result tables."}],"review_version":1}