{"id":"b9deeab6-3afb-4600-964a-4c31c93997e9","arxiv_id":"2509.08265","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HyMamba, a Mamba-based hyperspectral tracker that scans spatial and spectral dimensions and propagates a spectral hidden state across frames and layers, achieves state-of-the-art scores on seven HS tracking benchmarks.","lead":"A new video-tracking network, HyMamba, uses Mamba-style state space models to track objects in hyperspectral video, scanning both image directions and the spectral channel dimension and carrying a spectral memory across frames and network layers. It reports the best published accuracy on seven hyperspectral tracking benchmarks, for example 73.0% AUC on HOTC2020.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim hinges on unverified comparability of published baselines; the 0.3% AUC lead over SpectralTrack on HOTC2020 is within run-to-run noise.","rationale":"The paper's central claim is explicitly empirical: 'HyMamba achieves state-of-the-art performance' across seven benchmarks. The most direct way this claim could be false is if the comparison is not apples-to-apples. The authors do not state that baselines were re-run; they appear to reuse published numbers. Given the benchmark's maturity and the small margin over the closest competitor, this is the weakest link. The test-set hyperparameter selection compounds the issue for the flagship dataset. These concerns are not accusations; they are the standard reasons that empirical SOTA claims in tracking need code release and a clearly defined protocol. The architectural contribution is plausible: SSI/HSM modules are clearly described, ablations show consistent improvements, and the use of a strong frozen backbone (HiViT/SUTrack) is reasonable. But the exact SOTA margins are unverifiable without code and re-runs. Therefore the verdict remains CONDITIONAL: the paper should be accepted only if the authors release code and either re-run baselines or make a pre-registered validation split.","tokens_in":23265,"tokens_out":10124,"duration_ms":101745,"concrete_test":"Obtain the official HOTC2020 evaluation code and re-run the top HS trackers (especially SpectralTrack, SSTtrack, UBSTrack, SP-HST) under identical training splits and evaluation protocol; if the AUC gap between HyMamba and SpectralTrack narrows below 0.3% or reverses, the state-of-the-art claim for HOTC2020 fails. In parallel, retrain HyMamba with N and n chosen on a held-out validation split (e.g., leave-out videos from the 35-test set) and report the resulting AUC; if it no longer exceeds SpectralTrack's 0.727, the headline number is an artifact of test-set hyperparameter selection.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section IV-B compares HyMamba with 16 RGB and 16 HS trackers using numbers tabulated from their source publications. The paper states no re-runs under a unified protocol (same train/test splits, evaluation script, template-update policy, or hardware/software versions). The narrowest margin on the headline benchmark is 0.3% AUC (0.730 vs 0.727 for SpectralTrack); this gap is comparable to typical one-run variance in single-object tracking, so the SOTA claim is not robust without a controlled comparison. The same risk applies to Tables II-III, where many baselines are cited from concurrent papers. A second, related fragility is that the central hyperparameters (number of SSI layers N=4, spectral hidden state length n=16) are selected by maximizing AUC/DP@20 on the HOTC2020 test set (Table V, Fig. 8). Because the same test set is then used for the reported 0.730 result, that number is an optimistic in-sample estimate; it cannot be taken at face value as evidence of superiority until confirmed on a validation-selected configuration or on other datasets. The paper's internal ablations are informative but do not resolve these protocol questions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HyMamba, a Mamba/SSM-based tracker for hyperspectral video. It introduces a Spectral State Integration (SSI) module that propagates a recurrent spectral hidden state across transformer layers and frames, and a Hyperspectral Mamba (HSM) module with forward, backward, and spectral scanning SSMs. The network uses a frozen RGB-pretrained backbone and trains only the ASD module, SSI stack, and HS-specific patch embedding. Experiments on seven HS benchmarks report state-of-the-art results, e.g., 0.730 AUC and 0.963 DP@20 on HOTC2020, with ablations showing each proposed component contributes positively.","tokens_in":23575,"tokens_out":4715,"duration_ms":55350,"significance":"If verified, HyMamba would be a meaningful architectural contribution to hyperspectral tracking: it explicitly models cross-depth and temporal spectral information while avoiding the spectral-information loss of band selection/regrouping, and it demonstrates that Mamba-style state-space scanning can be adapted to the spectral dimension. Strengths include evaluation across seven benchmarks, consistent ablation results, an attribute-level analysis, and a commitment to release code. The central SOTA claim, however, currently rests on baseline numbers taken from heterogeneous publications and on hyperparameters selected on the same test set used for the headline results; the mathematical formulation of the SSM update also contains a dimension inconsistency. These issues are fixable but are load-bearing for the paper's main claim.","major_comments":[{"comment":"Equation (7) is dimensionally inconsistent as written. H_spec_i and H_spec_{i-1} are in R^n, F_HS_Linear is in R^{L×2D}, and B is declared as R^{n×1}; the term B F_HS_Linear cannot be computed, and even if broadcast, the result cannot be added to A H_{i-1} in R^n. The forward and backward SSM paths presumably have the same issue. Please provide the correct selective-scan formulation with explicit shapes for the input projection, the discretized A/B/C matrices, and the scan dimension. Without this, the central module is underspecified and not reproducible.","section":"III-C, Eqs. (6)-(7)"},{"comment":"The state-of-the-art claim compares HyMamba with 16 RGB and 16 HS trackers using numbers tabulated from their source publications. The paper does not state that baselines were re-run under a unified protocol (same train/test splits, evaluation script, template-update policy, and hardware/software versions). The narrowest margin on the headline benchmark is 0.3% AUC over SpectralTrack (0.730 vs 0.727), which is within typical run-to-run variation for single-object tracking. Without controlled re-runs or clearly stated protocols, the 'state-of-the-art' claim is not robust. Please provide a unified comparison or explicitly qualify the claim.","section":"IV-B, Tables I-III"},{"comment":"The number of SSI layers (N=4) and the spectral hidden-state length (n=16) are selected by maximizing AUC/DP@20 on HOTC2020, the same benchmark used for the reported headline result of 0.730 AUC. No held-out validation split is described. The reported number is therefore an optimistic in-sample estimate, and the margin over the closest competitor may be exaggerated. Please select hyperparameters on a validation set, or demonstrate that the chosen values are stable across the other six datasets.","section":"IV-E, Table V, Fig. 8"},{"comment":"No experiment in the paper reports multiple runs, standard deviations, or significance tests. Given the small margins over second-place methods on HOTC2020 (+0.3% AUC, +0.9% DP@20), the reported improvements may not be statistically meaningful. Please report variance across repeated runs, or at least confirm the key margins (especially against SpectralTrack, SP-HST, and MCITrack) with multiple seeds.","section":"IV-B, IV-E"}],"minor_comments":[{"comment":"The legend lists 'MCCT', 'SiamRPN++', and 'fDSST', which do not appear in Table I; 'MCCT' likely refers to MCITrack. Please harmonize tracker names and curve colors between the table and figures.","section":"Fig. 5"},{"comment":"The loss weights 5 and 2 in L_total are introduced without motivation or ablation. A brief justification or reference would help.","section":"Eq. (9)"},{"comment":"The dynamic template update is said to be controlled by a fixed temporal interval and a confidence threshold, but neither the interval nor the threshold value is reported in Section IV-A. Please provide these implementation details.","section":"III-A and IV-A"},{"comment":"The sentence 'for the HS feature, this step is only applied on F_HS_0' is ambiguous. It should be clarified whether the search filter is applied to the HS feature only in the first SSI layer, and why.","section":"III-C"},{"comment":"References [20] and [21] are given as URLs. If formal challenge papers or dataset descriptions exist, please cite them; otherwise, provide access dates and version information.","section":"References [20], [21]"}],"recommendation":"major_revision","confidential_remarks":"The paper's core architectural idea is interesting and the experimental scope is broad. The main risk is that the SOTA claim depends on comparing with published baseline numbers under non-identical protocols, and on tuning hyperparameters on the test set. I would ask the editor to require either a controlled re-run of the closest baselines or a tempering of the SOTA claim, plus a corrected mathematical formulation of Eq. (7), before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HyMamba is the first Mamba-based hyperspectral tracker and the architecture is more than a sticker: three-directional scanning (forward, backward, spectral) inside a spectral state integration module that carries hidden state across layers and frames is a real extension of Mamba. The paper describes the design clearly enough that replication is plausible, and the ablation tables (IV–VIII) support the internal logic. Each added module moves the numbers, and the spectral SSM is the largest contributor. The consistency of the gains across the seven datasets, including the 2023/2024 challenge sets, gives the core claim credibility.\n\nThe soft spot is where it usually is in this niche: evaluation protocol. N=4 and n=16 are chosen by scanning values on the HOTC2020 test set (Table V and Fig. 8), and then the same test set yields the headline 0.730 AUC. That is an in-sample optimist. And the 0.3% AUC edge over SpectralTrack on that benchmark is well within one-run noise; the paper never states that baselines were re-run, so the margins over other papers' published numbers are only as comparable as the underlying protocol. If you write off the exact ranking, the directional claim still holds—every component ablation is positive and the gains on other datasets are larger.\n\nThe other limitation is the missing code. Given that the paper says it will be released, this is a temporary issue, but it matters for a result whose headline margin is so thin. On balance, this is a serious empirical paper for hyperspectral tracking, not a breakthrough that changes the field. The new thing is the spectral-scanning SSM and the unification of cross-depth and temporal state. I'd send it to reviewers, but I'd ask them to make the authors clarify hyperparameter selection, report variance, and ideally re-run the closest baseline. After code release, I'd be happy to cite it as a leading method.","headline":"HyMamba is the first Mamba-based HS tracker and the internal ablation story is credible, but the headline SOTA margin on HOTC2020 is undercut by test-set hyperparameter selection and unverified baseline comparability.","tokens_in":24050,"tokens_out":3120,"would_cite":true,"duration_ms":30081,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims HyMamba, a hyperspectral tracker using state space modules over raw spectral bands, achieves state-of-the-art results on seven hyperspectral tracking benchmarks.","keywords":["hyperspectral object tracking","state space model","Mamba","spectral feature modeling","visual tracking","temporal context propagation","false-color compression","benchmark evaluation"],"falsifier":"Re-run the published baseline trackers on HOTC2020 with the same train/test splits, template sizes, and evaluation script; if SpectralTrack, SP-HST, or MCITrack reaches or exceeds HyMamba's 0.730 AUC / 0.963 DP@20 under identical conditions, the state-of-the-art claim is refuted. Also, remove the spectral SSM path from HSM; if AUC stays at 0.730 instead of dropping near the reported 1.0%, the spectral-scan mechanism is not causal.","tokens_in":1687,"feed_emoji":"🎯","tokens_out":2102,"duration_ms":81105,"temperature":0.7,"pith_summary":"This paper proposes HyMamba, a hyperspectral object tracker whose central claim is that tracking should model the full spectral dimension with state space modules, not compress hyperspectral frames into false-color RGB images. It reports that the HyMamba network, built from Spectral State Integration modules with a three-directional Hyperspectral Mamba scan, achieves state-of-the-art AUC and precision on seven hyperspectral tracking benchmarks, including 73.0% AUC and 96.3% DP@20 on HOTC2020. The design propagates a spectral hidden state across transformer depths and across video frames, so spectral information from earlier frames and layers contributes to the current prediction. If true, this means hyperspectral tracking can use the raw spectral band structure rather than hand-designed band selection or regrouping, which the paper argues lose inter-band correlations.","feed_headline":"Raw spectral bands beat false-color for HS tracking","feed_subtitle":"HyMamba's Mamba-style scans along wavelength and across frames hit 73.0 AUC on HOTC2020, beating prior HS trackers.","key_machinery":"Spectral State Integration (SSI): a module inserted between transformer encoder layers that carries a three-part recurrent spectral hidden state (forward, backward, and spectral scans) across depths and frames, plus Joint Augment and Spectral Augment cross-attention mechanisms that enrich both the joint false-color-plus-spectral feature and the raw hyperspectral feature. Its engine is Hyperspectral Mamba (HSM), which adapts the Mamba sequence model—linear-time, hidden-state scanning—by adding a spectral-channel SSM to the usual forward/backward spatial scanning, and fuses the three directional streams with a joint branch through element-wise products in a Multi-Directional Fusion Module. Thi","core_discovery":"HyMamba's central claim is that unifying spectral, cross-depth, and temporal modeling in one state-space network is the right way to track in hyperspectral video. Existing hyperspectral trackers either select or regroup bands into false-color images, or fuse modalities, losing fine-grained inter-band correlations and leaving layer-wise and frame-wise spectral semantics isolated. HyMamba instead feeds the unconverted hyperspectral cube through the Spectral State Integration module: a Hyperspectral Mamba scans features along forward, backward, and spectral-channel directions, updating a recurrent hidden state that passes from transformer layer to layer and from frame to frame, while joint and","pith_inferences":["The temporal hidden state invites a stress test on longer sequences with extended occlusions: if spectral continuity is doing the work, tracking should degrade gracefully with occlusion length, and if it degrades abruptly, the temporal propagation claim is weaker than stated.","A band-subsampling experiment could separate two effects the paper confounds: the number of spectral channels fed in and the architecture's ability to use them. Running the same network on random 8-, 16-, 25-, and 32-band subsets of the same videos would show where the benefit saturates.","The same three-directional scanning and cross-depth state propagation idea could carry over to other hyperspectral video tasks such as detection or segmentation, which also often compress spectral data or process frames independently."],"forward_implications":["If the central claim is right, hyperspectral tracking no longer needs to trade away spectral fidelity: feeding raw spectral bands into a state-space tracker beats converting them to one or several false-color images.","The recurrent spectral hidden state means information from earlier frames and shallower layers propagates into current predictions, so tracking should stay more stable through occlusion, fast motion, and deformation than per-frame trackers.","The same architecture works across 15-, 16-, and 25-band datasets, indicating the design transfers across spectral resolutions without per-dataset band selection.","Because only the ASD module, SSI/HSM modules, and hyperspectral patch-embedding layers are trained while the backbone stays frozen, the method can adapt to new hyperspectral benchmarks with limited training data.","The spectral SSM scan is a load-bearing component: the ablation reports that removing it costs 1.0% AUC and 1.7% DP@20 on HOTC2020."],"supporting_citations":[{"why":"Supplies the HOTC2020 benchmark, its CIE false-color conversion, and the MHT baseline; defines most of the evaluation datasets.","marker":"[1]"},{"why":"Defines the Mamba selective state space model that HSM adapts with forward, backward, and spectral scanning.","marker":"[15]"},{"why":"MCITrack is the strongest RGB-video competitor on HOTC2020 and provides the video-level contextual tracking baseline HyMamba outperforms.","marker":"[17]"},{"why":"SUTrack provides the pretrained initialization for the frozen backbone and tracking head, and is the ablation baseline.","marker":"[45]"},{"why":"HiViT is the frozen backbone whose transformer layers HyMamba inserts SSI modules into.","marker":"[48]"},{"why":"SpectralTrack is the closest prior hyperspectral tracker on HOTC2020 (0.727 AUC), the comparison that defines the paper's narrowest margin.","marker":"[60]"},{"why":"SSTtrack is the strongest competitor on the NIR2024 and RedNIR2024 benchmarks, used to establish state-of-the-art results there.","marker":"[64]"}],"fun_headline_variants":["Mamba-style scans capture spectral and temporal cues for HS tracking","Unified spectral, cross-depth, and temporal modeling in a single network","Raw spectral features beat false-color for hyperspectral video tracking","HyMamba leverages state-space models for cross-depth hyperspectral tracking","Track hyperspectral objects with Mamba's three-way scanning"],"cache_read_input_tokens":25856,"weakest_assumption_plain":"The load-bearing premise is that the published baseline scores on the seven benchmarks were measured under the same conditions as the authors' own runs; the smallest claimed win—0.3% AUC over SpectralTrack on HOTC2020—would be the first to disappear if any baseline changes under a re-run.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-style scans capture spectral and temporal cues for HS tracking","Unified spectral, cross-depth, and temporal modeling in a single network","Raw spectral features beat false-color for hyperspectral video tracking","HyMamba leverages state-space models for cross-depth hyperspectral tracking","Track hyperspectral objects with Mamba's three-way scanning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":1956,"prompt_tokens":784,"completion_tokens":1172,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1085}},"tokens_in":528,"tokens_out":1172,"duration_ms":11024,"temperature":1.0,"reasoning_tokens":1085,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:54:07.279055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the published baseline trackers on HOTC2020 with the same train/test splits, template sizes, and evaluation script; if SpectralTrack, SP-HST, or MCITrack reaches or exceeds HyMamba's 0.730 AUC / 0.963 DP@20 under identical conditions, the state-of-the-art claim is refuted. Also, remove the spectral SSM path from HSM; if AUC stays at 0.730 instead of dropping near the reported 1.0%, the spectral-scan mechanism is not causal.","supporting_citations":[],"review_version":1}