{"id":"d196638e-e33b-4333-a6c0-00986e843257","arxiv_id":"2505.04088","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SMMT, a Siamese tracker combining Motion Mamba and self-attention, reports improved precision and success on four TIR tracking benchmarks.","lead":"This paper presents a new infrared object tracker that combines a Mamba-style state-space model with self-attention inside a Siamese network. It reports higher accuracy than prior trackers on four thermal infrared benchmarks, but the method description has gaps and the hyperparameters were tuned on one of the evaluated benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed temporal motion modeling is not present in Eq. (2): only template and search features are transformed, with no temporal index or state, so the core novelty and mechanism behind the reported gains are unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the paper says the SMM extracts motion features from previous and next frame images, but the equations only process template and search features. I agree, and I think this is the most important flaw because the paper's central contribution is a motion-modeling architecture; if the temporal input is not defined, the ablation improvements cannot be explained by motion modeling. A conditional verdict is appropriate: the benchmark numbers, if reproducible, would still be evidence of empirical performance, but the mechanism and the architecture description need correction. The concern is addressable with a clarified dataflow or code release. I also note secondary issues, including alpha and beta being tuned on the PTB-TIR test set in Section 4.2 and the absence of a training/validation split or code, but those do not change the verdict and instead reinforce that the central empirical claim is not yet fully supported.","tokens_in":11206,"tokens_out":9354,"duration_ms":93203,"concrete_test":"Obtain the code or a precise dataflow for one forward pass and check whether any tensor is indexed by a frame time step or propagated as a state across consecutive frames. Specifically, trace each input to phi_V and phi_H in Eq. (2): if the only inputs are a template patch and a search patch from one frame pair, and the outputs are cross-correlated immediately, then the SMM performs spatial feature transformation, not motion modeling. Equivalently, ask the authors to annotate Fig. 2 with tensor shapes and frame indices; if no temporal dimension appears anywhere in the SMM, the 'motion modeling' claim should be revised or withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states that 'The SMM module extracts motion features from detection features of both the previous and next frame images,' and the PTB-TIR gains are attributed to this motion modeling. The formal definition in Section 3.3, Eq. (2), however, defines only F_V^Z = phi_SH(phi_V(F_base^Z)), F_H^Z = phi_SH(phi_H(F_base^Z)), F_V^X = phi_SH(phi_V(F_base^X)), and F_H^X = phi_SH(phi_H(F_base^X)). The inputs are template and search features F_base^Z and F_base^X; no previous/next-frame pair, temporal index, or cross-frame state appears. The HSS/VSS scans are spatial, along rows and columns, so even if the crops came from adjacent frames, this is not temporal motion modeling. In addition, phi_SH is described as the 'Siamese parameter-sharing strategy,' a weight-sharing constraint rather than a feature transformation, so Eq. (2) is not a well-formed operator, and the self-attention block mentioned in Section 3.2 is not represented there. This does not disprove the Table 2 numbers, but it removes the mechanistic basis of the central claim: the reported improvements cannot be attributed to the proposed motion-modeling architecture as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SMMT, a Siamese tracker for thermal infrared targets, which combines a ``Motion Mamba'' module with bidirectional horizontal/vertical state-space scans, a Siamese parameter-sharing strategy, and a motion edge-aware regression loss. The method is evaluated on LSOTB-TIR, PTB-TIR, VOT-TIR2015, and VOT-TIR2017, reporting strong results including accuracy scores of 0.79 and 0.74 and EAO scores of 0.376 and 0.345 on the VOT-TIR benchmarks, with component ablations on LSOTB-TIR and a parameter-sensitivity study on PTB-TIR. The central claim is that the proposed motion-modeling architecture and loss improve tracking accuracy and robustness in TIR imagery, especially under motion blur and occlusion.","tokens_in":11534,"tokens_out":5592,"duration_ms":52651,"significance":"If the architecture truly performed temporal motion modeling and the loss functions were as described, the reported gains would be a useful step for TIR tracking, a domain where motion blur, occlusion, and low texture are common. The paper has strengths: it evaluates on four established benchmarks, provides a component-wise ablation in Table 1, and reports scenario-level results on PTB-TIR. However, as written the manuscript does not provide a well-formed definition of the motion-modeling module, leaves several loss terms underspecified, and selects hyperparameters on PTB-TIR before reporting the final PTB-TIR results. These issues are load-bearing because they directly concern the core novelty and the validity of the reported benchmark comparisons. The mechanistic basis for the central claim is therefore not currently established by the text.","major_comments":[{"comment":"The claimed temporal motion modeling is not represented in the formal definition. Section 3.2 states that the SMM module extracts motion features from detection features of both the previous and next frame images, but Eq. (2) defines only transformations of the template feature F_base^Z and search feature F_base^X. The operators φ_V and φ_H are described as spatial row/column scans, and φ_SH is a Siamese parameter-sharing strategy rather than a feature transformation. There is no temporal index, no paired-frame input, and no cross-frame state in the equation, and the self-attention block mentioned in Section 3.2 is also absent. Consequently, the architecture as written does not implement temporal motion modeling, and the ablation gains in Table 1 cannot be attributed to that mechanism. The authors should either extend Eq. (2) with the actual temporal inputs and state-propagation equations, or substantially revise the motion-modeling claims.","section":"§3.2–3.3, Eq. (2)"},{"comment":"The loss functions are not fully specified. Equation (5) refers to gradients of ground-truth and predicted images, ∇I_i and ∇Î_i, but the tracker outputs bounding boxes and does not produce images; the binary mask M_i and the index ``A-th image'' are undefined. Equation (6) uses E_F defined in Eq. (7) as a cross-entropy between distributions p and q, yet p_i and q_i are never defined, and no explanation is given for how they are computed from the tracker's features or output. Equation (7) also introduces W_i, F_i, and F̂_i without defining their provenance or dimensions. These omissions make the loss non-reproducible and prevent verification of the stated contribution of the motion-edge and fine-grained losses. Please provide precise definitions for every variable, including shapes and where each quantity is computed.","section":"§3.4, Eqs. (4)–(7)"},{"comment":"The hyperparameters α and β in Eq. (4) are selected by scanning values on PTB-TIR, as described in Section 4.2, and then the final comparison in Section 4.2 reports PTB-TIR precision and success as the main result for that dataset. This is circular for PTB-TIR: the benchmark result is not an independent test of the loss coefficients. To support the reported PTB-TIR numbers, the authors should tune α and β on a held-out validation split and report PTB-TIR only as a test set, or justify the selection procedure and quantify its effect on the other benchmarks.","section":"§4.2, Ablation Study of Parameters"},{"comment":"The text claims ``the best EAO score of 0.345 on VOT-TIR2017,'' which is consistent with Table 2, but it also states that SMMT ``achieves a score of 0.376 on VOT-TIR2015,'' which is not the best EAO on VOT-TIR2015 because UDCT reports EAO 0.420 on that dataset. The sentence ``SMMT achieves a score of 0.376 on VOT-TIR2015 and the best EAO score of 0.345 on VOT-TIR2017'' should be corrected to avoid implying state-of-the-art performance on both benchmarks, and the ``superior performance'' wording should be qualified accordingly.","section":"Table 2, §4.2 Performance Comparison"}],"minor_comments":[{"comment":"Reference [24] is cited for the VOT-TIR2017 benchmark, but the bibliography entry is ``The visual object tracking vot2017 challenge results''; please cite the correct VOT-TIR2017 benchmark publication.","section":"§4.1"},{"comment":"The text refers to ``Table 3'' when listing experimental results, but no Table 3 appears in the manuscript; either include the missing table or correct the cross-reference.","section":"§4.2"},{"comment":"Reference [26] is cited as ``CFNet'' in the list of compared trackers, but the bibliography entry for [26] is for CEDNet; the reference and the cited method do not match.","section":"§4.2"},{"comment":"Figures 4 and 5 are used both in the parameter ablation (Figs. 4 and 5) and again in the performance comparison section, which also refers to Fig. 4 and Fig. 5; the figure numbering should be made consistent.","section":"Figures 4 and 5"},{"comment":"The phrase ``M_i is the binary mask for the A-th image'' contains a typo; it should be ``the i-th image.''","section":"Eq. (5)"},{"comment":"The ablation increments in Table 1 are reported without variance or statistical significance, and some differences are small (e.g., Siam-SMM vs Siam-SPS success 58.6 vs 59.0); please report multiple runs or indicate whether the differences are consistent across random seeds.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like an early draft: the equations do not match the prose in the central motion-modeling claim, the loss definitions are incomplete, and a cited table is missing. These issues, together with the circular hyperparameter selection on PTB-TIR, require substantial revision before the benchmark claims can be trusted. I am not recommending rejection because the core idea, if properly formalized and evaluated, could be a useful contribution to TIR tracking."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an incremental engineering combination—Mamba-style scans, self-attention, and a Siamese RPN baseline—for thermal infrared tracking, and if the numbers hold up it is a modest improvement. But the central mechanism as written does not exist. Section 3.2 says the SMM module extracts motion features from the previous and next frame images. Equation (2), the formal definition, takes only template features F_base^Z and search features F_base^X and applies row/column state-space scans. There is no temporal index, no frame pair, no state carried across frames. The stress-test note is accurate: this is spatial scanning, not temporal modeling. The self-attention block mentioned in Section 3.2 also does not appear in Eq. (2), so the two headline components are underspecified.\n\nWhat the paper does do: it assembles a plausible tracker and evaluates it on four standard benchmarks with an ablation table. That is real work, and the VOT/LSOTB comparisons are reasonably broad. The reported gains over SiamRPN++ and MDNet are within the range that a good engineering tweak could produce.\n\nThe soft spots are major. The loss functions in Eqs. (5)-(7) are not well-defined: predicted images, gradient operators, distributions p and q, masks M_i, and the alignment errors are never concretely specified, so the loss cannot be reproduced or checked. Alpha and beta are selected by scanning values on PTB-TIR and then the same dataset is used for the headline comparison, so the +3.4 precision number is partly a selection artifact. No code or trained models are released, which would be the natural way to confirm Table 2. The citation pattern is mostly fine; the issue is internal coherence, not reference stuffing.\n\nWho gets value: someone working on Mamba-based trackers might find the combination worth trying, but as written the paper should not be a citable reference. I would not send it to peer review yet. The right move is a major revision: make the motion modeling formal, actually use previous/next frames in the equations, define the losses, move hyperparameter selection to a validation split, and ideally release code. Until then the core claim is assertion rather than demonstrated mechanism.","headline":"The claimed temporal motion modeling never appears in the equations, so the paper's main mechanism is unsupported despite a familiar benchmark-style evaluation.","tokens_in":12089,"tokens_out":3334,"would_cite":false,"duration_ms":34995,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SMMT: a Siamese Motion Mamba tracker that reports top accuracy on thermal infrared tracking benchmarks.","keywords":["thermal infrared tracking","Siamese network","state-space model","Mamba","self-attention","motion modeling","parameter sharing","target tracking"],"falsifier":"Inspect the SMM module's forward pass or trace Eq. (2): if the module consumes only the template image $I_Z$ and the search-region image $I_X$, with no previous/next frame feature tensor, then no temporal motion modeling takes place. A decisive experiment is to feed the tracker only the current frame; if accuracy is unchanged, the motion-modeling claim is unsupported.","tokens_in":11021,"feed_emoji":"🎯","tokens_out":6695,"duration_ms":56464,"temperature":0.7,"pith_summary":"The paper claims that a Siamese tracker built around a Motion Mamba module — a bidirectional state-space scan with self-attention — plus a parameter-sharing strategy and a motion edge-aware regression loss is the new state of the art in thermal infrared (TIR) object tracking. On VOT-TIR2015 and VOT-TIR2017 it reports the highest accuracy of the compared trackers (0.79 and 0.74) and the best expected average overlap on VOT-TIR2017 (0.345). A sympathetic reader cares because TIR tracking must work under low contrast, occlusion, and motion blur, conditions that the paper argues are addressed by explicit motion modeling rather than appearance matching alone. The paper also reports large gains on LSOTB-TIR and PTB-TIR, with ablations attributing the improvement to the SMM module, the parameter-sharing strategy, and the new loss.","feed_headline":"Thermal tracker tops benchmarks with Motion Mamba","feed_subtitle":"SMMT reports best accuracy on VOT-TIR2015 and VOT-TIR2017 using state-space motion modeling.","key_machinery":"The central object is the Siamese Motion Mamba (SMM) module, a pair of directional state-space scan branches — a horizontal state-space (HSS) model and a vertical state-space (VSS) model — that scan the feature map row-wise and column-wise, followed by a multi-head self-attention block. It sits inside a SiamRPN++ style Siamese architecture with an adaptive scale-weighting mechanism that favors the 1/4, 1/8, or 1/16 feature scale based on target size. The other two pieces are the Siamese parameter-sharing strategy (SPS), which shares convolutional weights between the two scan directions in layers 3 and 4 of the backbone, and the motion edge-aware regression loss, which combines IoU and L1 box regression with a gradient-edge term and a fine-grained feature-alignment term.","core_discovery":"Put on its own terms, the paper claims that adding temporal motion modeling to a Siamese region-proposal tracker improves thermal infrared tracking. The Siamese Motion Mamba Tracker (SMMT) processes feature maps at three scales, scans them horizontally and vertically with a bidirectional state-space model, and applies multi-head self-attention at the final layer; a Siamese parameter-sharing strategy makes the two scan directions share convolutional weights in the deeper layers, and a motion edge-aware regression loss supervises both the box regression and the predicted edge/gradient field. The result is a tracker that the paper reports as best among all compared methods on four TIR benchmarks: highest accuracy 0.79 on VOT-TIR2015, 0.74 on VOT-TIR2017, best EAO 0.345 on VOT-TIR2017, and top precision/success on PTB-TIR and LSOTB-TIR.","pith_inferences":["Inference: the claimed temporal motion modeling stands or falls on whether previous and next frame features actually enter the SMM module; Eq. (2) only shows template and search-region features being processed, so the temporal path is the part most worth verifying.","Inference: if the gain is real, the same state-space scan plus edge-aware loss could transfer to other low-texture modalities such as depth, radar, or night-vision tracking, where edge detail is similarly scarce.","Inference: a natural cheap experiment is to ablate the self-attention block alone and the edge loss alone on a motion-blur-heavy subset; the paper's scenario-level PTB-TIR results suggest the edge loss carries most of the motion-blur improvement."],"forward_implications":["On LSOTB-TIR, replacing the SiamRPN++ features with the SMM module raises precision from 74.0 to 77.8 and success from 55.4 to 58.6; adding parameter sharing plus the loss (full SMMT) raises them to 85.4 and 62.8.","On the VOT-TIR benchmarks, SMMT reports the best accuracy of every compared tracker on both VOT-TIR2015 (0.79) and VOT-TIR2017 (0.74), and the highest EAO on VOT-TIR2017 (0.345).","On PTB-TIR, the tracker reaches 83.9% precision and 63.6% success, and the paper reports it as the best method in the occlusion, fast-motion, and motion-blur scenario splits.","Because the parameter-sharing strategy reduces redundant computation between scan directions, the paper's design implies motion modeling can be added to Siamese trackers without multiplying the parameter count."],"supporting_citations":[{"why":"SiamRPN++, the Siamese region-proposal baseline that SMMT extends and that all ablations compare against.","marker":"[3]"},{"why":"Mamba, the selective state-space sequence model whose horizontal and vertical scans form the core of the SMM module.","marker":"[16]"},{"why":"MM-Tracker, the Motion Mamba design with adaptive scale selection that the SMM module adapts.","marker":"[17]"},{"why":"MambaVision, the hybrid state-space/transformer architecture cited as inspiration for adding self-attention blocks.","marker":"[18]"},{"why":"The backbone-splitting parameter-sharing strategy that motivates the Siamese parameter-sharing strategy (SPS).","marker":"[19]"},{"why":"PTB-TIR, the pedestrian TIR benchmark used for the precision/success and scenario-level comparisons.","marker":"[21]"},{"why":"LSOTB-TIR, the large-scale benchmark used for the main ablation study and final comparison.","marker":"[22]"},{"why":"VOT-TIR2015 challenge benchmark where SMMT reports the best accuracy of compared trackers.","marker":"[23]"},{"why":"VOT-TIR2017 challenge benchmark where SMMT reports the highest accuracy and EAO.","marker":"[24]"}],"fun_headline_variants":["Motion Mamba tracker beats all on four thermal benchmarks","Siamese Motion Mamba tops TIR benchmarks with state-space motion","Bidirectional Mamba and self-attention boost thermal tracking","SMMT: motion-aware regression improves thermal object tracking","Thermal tracking gets a Mamba boost in Siamese architecture"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Motion Mamba module actually receives temporal information from previous and next frames; the paper's equations define only template and search-region features, so if that temporal input path is absent the claimed motion modeling would not occur.","fun_headline_variants_meta":{"raw":{"variants":["Motion Mamba tracker beats all on four thermal benchmarks","Siamese Motion Mamba tops TIR benchmarks with state-space motion","Bidirectional Mamba and self-attention boost thermal tracking","SMMT: motion-aware regression improves thermal object tracking","Thermal tracking gets a Mamba boost in Siamese architecture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2366,"prompt_tokens":928,"completion_tokens":1438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1355}},"tokens_in":544,"tokens_out":1438,"duration_ms":11577,"temperature":1.0,"reasoning_tokens":1355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:37:22.642245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the SMM module's forward pass or trace Eq. (2): if the module consumes only the template image $I_Z$ and the search-region image $I_X$, with no previous/next frame feature tensor, then no temporal motion modeling takes place. A decisive experiment is to feed the tracker only the current frame; if accuracy is unchanged, the motion-modeling claim is unsupported.","supporting_citations":[{"cited_title":"In: ICCV, 4282-4291 (2019)","cited_arxiv_id":null,"evidence_quote":"SiamRPN++, the Siamese region-proposal baseline that SMMT extends and that all ablations compare against."},{"cited_title":"MM-Tracker: Motion Mamba with Margin Loss for UAV-platform Multiple Object Tracking","cited_arxiv_id":"2407.10485","evidence_quote":"MM-Tracker, the Motion Mamba design with adaptive scale selection that the SMM module adapts."},{"cited_title":": Parameter sharing and multi -granularity feature learning for cross-modality person re-identification","cited_arxiv_id":null,"evidence_quote":"The backbone-splitting parameter-sharing strategy that motivates the Siamese parameter-sharing strategy (SPS)."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"PTB-TIR, the pedestrian TIR benchmark used for the precision/success and scenario-level comparisons."},{"cited_title":"In: ACM MM, 3847-3856 (2020)","cited_arxiv_id":null,"evidence_quote":"LSOTB-TIR, the large-scale benchmark used for the main ablation study and final comparison."},{"cited_title":"In: ICCV Workshops, 76-88 (2015)","cited_arxiv_id":null,"evidence_quote":"VOT-TIR2015 challenge benchmark where SMMT reports the best accuracy of compared trackers."},{"cited_title":"In: ICCV Workshops, 1949–1972 (2017)","cited_arxiv_id":null,"evidence_quote":"VOT-TIR2017 challenge benchmark where SMMT reports the highest accuracy and EAO."}],"review_version":1}