{"id":"ecd16cb0-97b0-4250-af8c-95aff568f1c7","arxiv_id":"2506.03335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Mamba-plus-attention motion predictor with a height-adaptive IoU matching metric achieves state-of-the-art HOTA on SportsMOT and strong zero-shot results on VIP-HTD.","lead":"SportMamba is a new multi-object tracking system for team sports that combines a Mamba state-space motion predictor with self-attention and a height-aware matching metric to reduce identity switches. On the SportsMOT benchmark it reaches the best reported HOTA score (77.3) and transfers to ice hockey without retraining.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Detector training split is the load-bearing assumption: SportMamba is unmarked in the train+val asterisk convention but the paper never states on which split its YOLOX was trained, so the +0.1 HOTA SOTA margin may be unsupported.","rationale":"The reader's weakest assumption centers on detector split and unreported hyperparameters; my independent read converges on the same point. The comparative claim is the central claim, and the table's asterisk convention makes the detector split decisive. The other issues raised in the paper — the questionable HIoU definition (Eq. 8), the lack of an MHSA-only ablation, and missing code — are real but secondary. The HIoU absolute value is multiplied by EIoU, so even if the formula misbehaves for vertically disjoint boxes, the final association score is still zero when the areas do not overlap; it therefore does not by itself overturn the empirical comparison. The MHSA ablation affects attribution of the improvement, not the headline SOTA. The missing code matters for reproducibility, but the immediate logical keystone is the training-data protocol. Because the ambiguity is explicit and easily resolved by a config release, the paper remains conditional rather than rejected: there is no evidence that the authors acted improperly, only that the required condition for the SOTA claim has not yet been established.","tokens_in":12997,"tokens_out":7669,"duration_ms":86746,"concrete_test":"Ask the authors for the full detector and tracker config (YOLOX training split, ReID checkpoint, λ_reid, λ_ssim, α, σ) or release code. Then rerun the SportsMOT evaluation with a YOLOX detector trained exclusively on the SportsMOT training split, keeping the tracker settings fixed, and recompute HOTA, IDF1, AssA, and DetA. If SportMamba still exceeds Deep-EIoU in HOTA and the table values are reproduced, the SOTA claim is supported. If HOTA drops below 77.2 or the margin reverses, the claimed SOTA depends on the unstated detector training data and should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result is a benchmark comparison: SportMamba reaches HOTA 77.3 on SportsMOT and HOTA 65.1 zero-shot on VIP-HTD, beating prior trackers. For that claim to be true, the comparison must be apples-to-apples. Table 1 encodes this in a * convention: methods marked with * train the detector on train+val; SportMamba is unmarked, which asserts train-only. However, §4.1 only says 'Following prior works, we train a YOLOX model' and never states the split used. The same omission affects the VIP-HTD 'zero-shot' experiment, which requires that the detector and motion predictor see no VIP-HTD training data. The association hyperparameters are also absent: λ_reid and λ_ssim in Eq. (10), α and σ in Eq. (12), and the ReID backbone are unspecified. Since the HOTA margin over Deep-EIoU is only +0.1 (77.3 vs 77.2) and DetA is 1.3 higher, a detector trained on more data could plausibly account for the difference. This is a verification gap, not evidence of misconduct; it is the load-bearing assumption because if the detector protocol is not train-only, the SOTA claim falls regardless of the motion predictor and association components.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SportMamba, a tracking-by-detection multi-object tracker for team sports. The method consists of a YOLOX detector, a Mamba-based motion predictor augmented with multi-head self-attention, a hybrid association cost that combines ReID appearance features with a height-adaptive extended IoU (HA-EIoU), and a two-stage high/low confidence association pipeline with dynamic EMA feature updating. The authors report state-of-the-art HOTA on the SportsMOT test set (77.3) and the highest HOTA on VIP-HTD in a zero-shot setting (65.1), with ablations on SportsMOT validation showing that HA-EIoU improves over EIoU and that buffer sizes and Mamba block counts affect performance.","tokens_in":13281,"tokens_out":6480,"duration_ms":73699,"significance":"If the results are reproducible under the stated protocol, the paper demonstrates a meaningful improvement on a challenging sports MOT benchmark: the Table 1 HOTA margin over Deep-EIoU is small (77.3 vs 77.2) but positive, and the zero-shot VIP-HTD result in Table 2 is a useful generalization datapoint. The validation ablation in Table 5 cleanly isolates the association metric and gives quantitative support for the HA-EIoU idea, and Tables 3 and 4 provide systematic sweeps over M, w, and buffer sizes. However, the significance is currently limited by missing protocol details: the detector training split is not stated, several association hyperparameters and the ReID backbone are unreported, the MHSA component is not ablated, and the HIoU formula in Eq. (8) appears to contain a sign error. These gaps prevent the reader from verifying that the headline comparison is apples-to-apples.","major_comments":[{"comment":"The training split for the YOLOX detector is never stated. Table 1 uses an asterisk to mark methods that train the detector on train+val, and SportMamba is unmarked, which implicitly asserts train-only training; however, §4.1 only says 'Following prior works ..., we train a YOLOX model' without specifying the split. Because the HOTA margin over Deep-EIoU is only 0.1 (77.3 vs 77.2), and because the same omission affects the zero-shot VIP-HTD experiment in Table 2, the SOTA claim is not verifiable unless the authors state the exact split and confirm that no VIP-HTD training data were used for the detector or motion predictor.","section":"§4.1, Tables 1–2"},{"comment":"The absolute value in the HIoU formula is incorrect for non-overlapping bounding boxes. When min(y2, y-hat2) - max(y1, y-hat1) is negative, the absolute value converts the vertical gap into a positive numerator, so vertically separated boxes receive a positive height-similarity score that grows with separation instead of a score of zero for no overlap. Since HA-EIoU is a central contribution and is the basis of the Table 5 ablation, the formula should be corrected to use a clamped overlap, e.g., max(0, min(y2, y-hat2) - max(y1, y-hat1)) divided by the union height, and the experiments should be re-run or the corrected metric should be justified.","section":"Eq. (8), §3.3.1"},{"comment":"The claimed benefit of the Mamba-attention mechanism is not ablated. Table 3 varies the number of Mamba-attention blocks M and the tracklet window w, but never removes the MHSA block to compare against a vanilla Mamba encoder. Without such a comparison, the paper's first stated technical contribution, the mamba-attention motion predictor, is not empirically supported.","section":"§3.2, Table 3"},{"comment":"Several load-bearing hyperparameters are not reported: the association weights lambda_reid and lambda_ssim in Eq. (10), the EMA smoothing base alpha and the minimum confidence threshold sigma in Eq. (12), and the backbone used for ReID feature extraction. These values are needed to reproduce the method and to rule out that they were selected on the test set. The hybrid cost and dynamic EMA are central pipeline components, so this omission is a reproducibility gap.","section":"Eqs. (10), (12), §4.1"},{"comment":"The buffers b1 and b2 are described as 'adaptive' in the abstract and contributions, but in the method they are fixed scalar hyperparameters selected by the validation sweep in Table 4; no per-detection, per-frame, or per-sequence adaptation mechanism is described. Either the mechanism should be made adaptive, or the terminology should be revised to avoid overclaiming.","section":"Abstract and §3.3"}],"minor_comments":[{"comment":"The dataset name is misspelled as 'sportMOT' in the conclusion; it should be 'SportsMOT'.","section":"§5"},{"comment":"The dataset name 'VIT-HTD' should be 'VIP-HTD'.","section":"Figure 6 caption"},{"comment":"References [39] and [40] appear to be the same HockeyMOT paper, and references [43] and [44] are the same MambaTrack paper; duplicate entries should be merged.","section":"References"},{"comment":"The smooth L1 loss is missing the conventional 0.5 factor in the quadratic region; either correct the formula or state the chosen convention explicitly.","section":"Eq. (14)"},{"comment":"The dynamic EMA update is undefined for detection confidences below sigma, since s_t - sigma can be negative and alpha_d can fall outside a meaningful range; the allowed domain of s_t for this update should be stated.","section":"Eq. (12)"},{"comment":"The validation HOTA scores in Tables 3–5 are near 84 while the test HOTA in Table 1 is 77.3; this large gap is not discussed and a brief explanation would help readers interpret the ablations.","section":"Tables 3–5 vs Table 1"},{"comment":"The abstract claims 'state-of-the-art performance on various metrics,' but Table 1 shows that Deep-EIoU has higher IDF1 (79.8 vs 77.7) and AssA (67.7 vs 66.8); the claim should be qualified to HOTA and DetA.","section":"Abstract and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper does not state whether code or models will be released, which matters for a result whose headline margin is 0.1 HOTA. The overlap with Deep-EIoU, CBIoU, and MambaTrack is substantial, and the current manuscript does not fully disentangle which component produces the improvement; the missing MHSA ablation and the detector-split ambiguity are the key issues to resolve. The duplicate references and small typos suggest the manuscript was assembled quickly and needs a careful revision pass."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent piece of MOT engineering for team sports, and the core ingredients are real: combining Mamba with multi-head self-attention for motion prediction, and a height-adaptive IoU with expanded buffers for association. The reported validation ablation shows HA-EIoU is worth about 0.9 HOTA over plain EIoU, which is a genuine, if modest, gain. The zero-shot transfer to VIP-HTD is a useful sanity check, and the 30 FPS claim is sensible for a Mamba-based tracker. On the evidence in the paper, the headline result — state of the art on SportsMOT at 77.3 HOTA — is plausible but not fully verified.\n\nThe biggest soft spot is exactly what the stress test flags: the detector training split is never stated. Table 1 uses the * convention to mark methods that train on train+val, and SportMamba is unmarked, which implicitly claims a train-only detector. But Section 4.1 just says \"we train a YOLOX model\" without saying on which split. Since the margin over Deep-EIoU is only 0.1 HOTA, an unstated train+val detector could account for the difference. This is a verification gap, not evidence of misconduct, but it is the load-bearing assumption behind the SOTA claim.\n\nTwo smaller issues. First, the paper never ablates the MHSA addition to the Mamba encoder, so the motion-predictor contribution is not isolated. Second, the HIoU formula in Eq. (8) takes an absolute value of the numerator, which makes it positive for vertically separated boxes that do not overlap at all, so it can reward a large vertical gap. In practice the metric may work because other terms dominate, but the definition is wrong as written. Hyperparameters like lambda_reid, lambda_ssim, alpha, sigma, and the ReID backbone are also missing, and no code is provided.\n\nNone of these are fatal. The ablations support the association-metric improvement, the zero-shot result is a reasonable generalization check, and the writing is clear. The paper is exactly the kind of work a serious referee should see: the claims are specific and benchmarkable, and the missing details are addressable in revision. I would send it to peer review, but I would not cite it as a definitive SOTA comparison until the detector split and code or detailed configs are public.","headline":"Solid engineering contribution with a plausible but under-specified SOTA claim; the detector split question is the load-bearing issue.","tokens_in":13859,"tokens_out":1703,"would_cite":false,"duration_ms":18744,"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":"This paper claims that a Mamba-attention motion predictor combined with a height-adaptive IoU association metric achieves state-of-the-art multi-object tracking on team-sports datasets.","keywords":["multi-object tracking","Mamba","state space models","motion prediction","data association","occlusion handling","team sports","SportsMOT"],"falsifier":"Retrain and rerun SportMamba on SportsMOT with the detector trained strictly on the train split, with $\\lambda_{\\mathrm{reid}}$ and $\\lambda_{\\mathrm{ssim}}$ chosen by validation, and compare HOTA against Deep-EIoU under the same detector and ReID settings; if the reported margin disappears or reverses, the claimed advantage is not supported.","tokens_in":12756,"feed_emoji":"⚽","tokens_out":6372,"duration_ms":69759,"temperature":0.7,"pith_summary":"SportMamba claims that the two difficulties that break trackers in team sports—nonlinear player motion and identity switches under occlusion—can be addressed together by a hybrid tracking-by-detection pipeline. It couples a motion predictor that adds self-attention on top of a Mamba state-space model to capture non-linear trajectories, with a height-adaptive spatial association metric that weights bounding-box overlap by vertical alignment before matching. On the SportsMOT test set the paper reports HOTA 77.3, IDF1 77.7, and DetA 89.5, ahead of prior learning-based trackers, and on VIP-HTD ice hockey it reports HOTA 65.1 without any tracker-specific fine-tuning. The result matters because it suggests a practical, near-real-time (about 30 FPS) recipe for sports tracking that does not rely heavily on appearance cues, which are unreliable when teammates wear identical jerseys.","feed_headline":"State-space tracker hits 77.3 HOTA on SportsMOT","feed_subtitle":"A Mamba-attention predictor plus a height-aware IoU keeps player identities stable through fast motion and occlusion.","key_machinery":"The load-bearing mechanism is the pairing of a Mamba-attention motion predictor with a height-adaptive spatial association cost. The motion predictor embeds a tracklet of past bounding boxes, passes them through $M$ stacked blocks that alternate a selective state-space model (Mamba) with multi-head self-attention, and regresses the next-frame box with an MLP. The association side replaces the standard IoU with $\\mathrm{HA\\text{-}EIoU} = \\mathrm{HIoU} \\cdot \\mathrm{EIoU}$, where EIoU computes IoU on boxes widened by buffers $b_1$ and $b_2$, and HIoU is the ratio of vertical-overlap height to total vertical extent, so matches require both horizontal overlap and depth-consistent height alignment. A weighted sum with appearance cosine similarity feeds a Hungarian solver, and unmatched low-confidence detections are recovered in a second pass using only HA-EIoU. Adaptive buffer sizes and a confidence-dependent EMA feature update complete the pipeline.","core_discovery":"The paper's central claim is that replacing the usual linear-motion assumption with a learned Mamba-attention motion predictor, and replacing plain IoU with a height-adaptive extended IoU, is enough to push multi-object tracking in team sports past previous methods. Concretely, SportMamba scores 77.3 HOTA, 77.7 IDF1, and 89.5 DetA on the SportsMOT test set, the highest HOTA among the compared methods, and 65.1 HOTA on VIP-HTD in a zero-shot setting where no tracker-specific training on that dataset is performed. The paper attributes the gains to two components working together: the motion predictor models non-linear, non-appearance-based motion, and the height-adaptive metric (HA-EIoU) reduces wrong matches when players partially occlude one another and their box heights change with depth. The qualitative results show identity preservation after occlusion where the diffusion- and Mamba-based baselines fragment tracks.","pith_inferences":["A testable extension the paper leaves implicit is to apply the HA-EIoU cost to filter-based trackers such as ByteTrack or OC-SORT while keeping their Kalman motion model; if most of the gain transfers, the contribution lives in the association metric rather than in the Mamba predictor.","The adaptive buffers $b_1$ and $b_2$ are chosen by a validation sweep rather than learned; one could make them functions of predicted speed or detection confidence so the search space grows only when motion is fast, a natural next step not explored here.","To the extent that the detector is trained on SportsMOT and then applied to VIP-HTD, the zero-shot evidence tests the motion predictor and association metric more than the detector itself, so the generalization claim should be read with that scope in mind."],"forward_implications":["On the SportsMOT test set the method reports the highest HOTA among the compared trackers (77.3), including filter-based and learning-based baselines, at roughly 30 FPS inference.","The two-stage association means low-confidence detections are still consumed, so the tracker can recover tracks that momentarily lose high-confidence detections, which is the regime where identity switches occur in sports footage.","The zero-shot VIP-HTD result (HOTA 65.1) indicates that the learned motion predictor and height-adaptive metric transfer to a faster sport without retraining.","Height adaptation alone contributes about 0.9 HOTA and 1.1 IDF1 over extended IoU on the SportsMOT validation set, according to the ablation, so depth-scale alignment is a measurable source of the gain."],"supporting_citations":[{"why":"Supplies the selective state-space sequence model at the core of the motion predictor.","marker":"[19]"},{"why":"Provides the SportsMOT dataset and baseline results that define the main benchmark.","marker":"[8]"},{"why":"Introduces the extended-IoU and deep-features sports association baseline that the height-adaptive metric extends and is compared against.","marker":"[24]"},{"why":"Supplies the two-stage high/low-confidence association strategy and detection thresholding that SportMamba adapts.","marker":"[52]"},{"why":"Motivates expanding the matching space with buffers for irregular fast motion.","marker":"[47]"},{"why":"Diffusion-based non-linear motion predictor baseline that SportMamba must outperform.","marker":"[29]"},{"why":"Vanilla-Mamba motion prediction baseline whose hidden-state design SportMamba extends with self-attention.","marker":"[43]"},{"why":"Provides the VIP-HTD ice hockey benchmark used for the zero-shot generalization test.","marker":"[32]"},{"why":"YOLOX detector that SportMamba fine-tunes for player detection.","marker":"[14]"}],"fun_headline_variants":["Mamba motion model beats prior trackers in team sports","Non-linear tracking with Mamba attention scores 77.3 HOTA","Height-adaptive IoU reduces ID switches in occluded team play","Zero-shot transfer to hockey proves generalized tracking","Learned non-linear motion keeps player IDs stable through occlusions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that SportMamba is state of the art assumes the comparison is apples-to-apples: the paper does not report whether the YOLOX detector was trained only on the SportsMOT training split, and it does not give the association weights $\\lambda_{\\mathrm{reid}}$ and $\\lambda_{\\mathrm{ssim}}$ or the ReID model, so the gap over prior methods could shrink if those were fixed differently.","fun_headline_variants_meta":{"raw":{"variants":["Mamba motion model beats prior trackers in team sports","Non-linear tracking with Mamba attention scores 77.3 HOTA","Height-adaptive IoU reduces ID switches in occluded team play","Zero-shot transfer to hockey proves generalized tracking","Learned non-linear motion keeps player IDs stable through occlusions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2707,"prompt_tokens":973,"completion_tokens":1734,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1650}},"tokens_in":589,"tokens_out":1734,"duration_ms":15223,"temperature":1.0,"reasoning_tokens":1650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:05:26.655877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain and rerun SportMamba on SportsMOT with the detector trained strictly on the train split, with $\\lambda_{\\mathrm{reid}}$ and $\\lambda_{\\mathrm{ssim}}$ chosen by validation, and compare HOTA against Deep-EIoU under the same detector and ReID settings; if the reported margin disappears or reverses, the claimed advantage is not supported.","supporting_citations":[{"cited_title":"Sportsmot: A large multi- object tracking dataset in multiple sports scenes","cited_arxiv_id":null,"evidence_quote":"Provides the SportsMOT dataset and baseline results that define the main benchmark."},{"cited_title":"Iterative scale-up expansioniou and deep features association for multi-object tracking in sports","cited_arxiv_id":null,"evidence_quote":"Introduces the extended-IoU and deep-features sports association baseline that the height-adaptive metric extends and is compared against."},{"cited_title":"Bytetrack: Multi-object tracking by associating every detection box","cited_arxiv_id":null,"evidence_quote":"Supplies the two-stage high/low-confidence association strategy and detection thresholding that SportMamba adapts."},{"cited_title":"Hard to track objects with irregular motions and sim- ilar appearances? make it easier by buffering the matching space","cited_arxiv_id":null,"evidence_quote":"Motivates expanding the matching space with buffers for irregular fast motion."},{"cited_title":"Diffmot: A real-time diffusion-based multiple object tracker with non-linear prediction","cited_arxiv_id":null,"evidence_quote":"Diffusion-based non-linear motion predictor baseline that SportMamba must outperform."},{"cited_title":"Vip-htd: A public benchmark for multi-player tracking in ice hockey.Journal of Computa- tional Vision and Imaging Systems, 9(1):22–25, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the VIP-HTD ice hockey benchmark used for the zero-shot generalization test."}],"review_version":1}