{"id":"0a38d07a-6ada-4c48-ab64-3d01c4b23946","arxiv_id":"2507.20763","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"KASportsFormer combines bone and limb tokens with cross-attention in a spatio-temporal transformer, reporting state-of-the-art MPJPE on SportsPose and WorldPose short-video benchmarks.","lead":"A new transformer architecture for 3D human pose estimation from short sports videos adds kinematic bone and limb features to a spatio-temporal model. It reports small accuracy gains over prior methods on two sports datasets, but the improvements rest on custom evaluation protocols and lack statistical validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WorldPose's per-game clip split can leak scene/temporal context between train and test; without a held-out-game or official split, the 1.4 mm MPJPE SOTA margin is not established.","rationale":"The reader's conditional verdict is well-founded. I read the paper in good faith: the method is a reasonable engineering contribution that combines bone/limb tokens with cross-attention in a MotionAGFormer-like transformer; ablations show each component helps, and all baselines are retrained on the same data, so within-protocol comparisons are controlled. However, the claim that matters — 'state-of-the-art on SportsPose and WorldPose' — is only as strong as the evaluation protocol. The WorldPose split is the weakest link: saying clips of 'each games' are randomly categorized into train/test means the same game contributes to both sets, creating scene and temporal correlation that can inflate results. The SportsPose split is also custom (7 arbitrary subjects) and not justified against the official benchmark. The reported margins (0.1–1.5 mm) are small, and with single runs they are plausibly noise. A single held-out-game or official-split experiment would settle whether the claimed SOTA generalizes. I therefore agree with the reader's weakest assumption and keep the verdict conditional.","tokens_in":16433,"tokens_out":6276,"duration_ms":69845,"concrete_test":"Re-run the WorldPose experiments using a game-disjoint split (e.g., train on 6 games, test on the remaining 2 games, or use the official WorldPose train/test split if released) and the SportsPose official subject split, retraining KASportsFormer and the same baselines under identical settings; if KASportsFormer's MPJPE advantage over MotionAGFormer-B/L drops below ~1 mm or reverses, the SOTA claim does not survive the leakage-free protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim hinges on the evaluation protocol. In Section 4.1, WorldPose clips are 'randomly categorize[d] ... of each games into training and testing sets', so clips from the same soccer game (same broadcast camera setup, players, field, and possibly overlapping timestamps) appear on both sides of the train/test boundary. This allows the model to exploit scene- or player-specific cues rather than generalize to new games, which is exactly what the SOTA claim requires. The reported 1.4 mm MPJPE improvement over MotionAGFormer-B and 0.8 mm over D3DP (Table 3) could shrink or reverse under a game-disjoint split. The SportsPose protocol is also non-standard: the paper randomly selects 7 of 24 subjects for testing without showing this matches the official SportsPose benchmark split; since all baselines are retrained on the same custom split, the internal comparison is controlled, but the label 'state-of-the-art' does not extend to the official benchmark. With single runs and no error bars, margins of 0.1–1.5 mm are within typical run-to-run variance, so the split protocol is the load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes KASportsFormer, a 2D-to-3D lifting transformer for short (27-frame) monocular sports video. It augments the joint coordinate stream with two hand-designed anatomy streams: BoneExt converts a frame into bone direction and length tokens (including an averaged 'ultra-bone'), and LimbFus composes bones into biological and imaginary 'hyper-limb' tokens through per-limb MLPs. An anatomy mixer performs spatial and temporal multi-head cross-attention using bone queries and limb keys/values, in parallel with a MotionAGFormer-style joint MHSA/GCN mixer, followed by adaptive blending. The model is trained with per-joint position and velocity losses. All baselines are retrained on the same 27-frame custom splits; the authors report MPJPE/P-MPJPE improvements on SportsPose (58.0 mm DET MPJPE) and WorldPose (34.2 mm DET MPJPE) and claim state-of-the-art performance. Code and models are promised at a GitHub repository.","tokens_in":16637,"tokens_out":14761,"duration_ms":153335,"significance":"Strengths: the proposed bone and limb features are deterministic functions of the input coordinates, and all trainable components are supervised by ground-truth pose and velocity losses, so there is no circularity in the contribution claim. The experimental setup is internally consistent: every baseline is retrained on the same 27-frame protocol, and the ablations in Table 5 show that each module (BoneExt, LimbFus, MHCA) contributes to the reported result. Release of code and models is a concrete reproducibility asset. If the protocol concerns are resolved, the paper is a solid incremental architecture contribution that may be useful for sports-analysis applications. However, the headline 'state-of-the-art' claim currently rests on custom and potentially leaky splits, on single runs, and on margins as small as 0.1 mm, so the benchmark-level significance is not yet established.","major_comments":[{"comment":"The WorldPose evaluation protocol is the load-bearing assumption behind the SOTA claim. The text says clips of each game are 'randomly categorize[d]' into training and testing sets; because clips from the same broadcast game share camera calibration, players, field appearance, and possibly temporally adjacent frames, this per-clip split can leak game-specific cues into training. The reported DET MPJPE gain of 1.4 mm over MotionAGFormer-B and GT gain of 0.8 mm over MotionBERT (Table 3) could therefore reflect scene/player memorization rather than generalization to unseen games. Please evaluate on held-out games (or the official WorldPose protocol), clarify the ambiguous '26 clip sources' sentence, and report whether any training and testing clips are temporally overlapping. Unless the held-out-game numbers show the same margin, the Section 4.3 SOTA statement should be restricted to the custom split.","section":"Section 4.1 (WorldPose), Table 3"},{"comment":"The SportsPose evaluation also uses a custom protocol: the paper says the dataset is 'randomly divide[n]' into a 7-subject test set, without demonstrating that this matches the official SportsPose benchmark split. Because all baselines are retrained on the same split, the internal ranking is controlled. However, the claim that KASportsFormer 'achieved the current state-of-the-art (SOTA) performance on 27 frame video clips' implies comparability to published benchmark results, which is not valid unless the official evaluation protocol is used. Please either adopt the official SportsPose split or explicitly frame Table 1 as results on a custom protocol, and avoid the SOTA label in that case.","section":"Section 4.1 (SportsPose), Table 1"},{"comment":"All results are single runs, with no standard deviation or significance testing. Improvements over the next-best method are as small as 0.1 mm (SportsPose GT P-MPJPE relative to D3DP, Table 1) and 0.2-0.4 mm on several per-action rows in Table 2. Without repeated seeds, these margins are within typical run-to-run variance for transformer pose-lifting models. Please report mean and standard deviation over at least three seeds for the main comparisons, or otherwise temper the claims that depend on these small margins.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The abstract reports WorldPose MPJPE as 34.3 mm, but Table 3 reports 34.2 mm; please reconcile.","section":"Abstract / Table 3"},{"comment":"The sentence 'we outperformed D3DP by 0.8mm MPJPE and 0.4mm P-MPJPE' does not match Table 3: the 0.8 mm GT MPJPE gap is relative to MotionBERT, the GT P-MPJPE gap to MotionBERT is 0.6 mm, while the gaps to D3DP are 10.3 mm and 2.7 mm. Please correct the text or the table.","section":"Section 4.3 (WorldPose)"},{"comment":"The caption says 'Qualitative comparison' but the table is quantitative; change it to 'Quantitative comparison'.","section":"Table 3 caption"},{"comment":"The term 'descendant' is not clearly defined; for a directed bone from a joint toward the root, the closer node is usually the parent or ancestor. Please state the bone-direction convention explicitly.","section":"Section 3.2, Eq. (3)"},{"comment":"The text defines H^i_AC, H^i_AS, and H^i_G as the outputs of layer i, but Eq. (9) blends H^{i-1}_AC, H^{i-1}_AS, and H^{i-1}_G; clarify whether the blending uses current or previous layer outputs.","section":"Section 3.3, Eq. (9)"},{"comment":"The caption says 'number of hidden layers' while the text describes h_id as the hidden layer dimension; these are different quantities and should be made consistent.","section":"Table 4"},{"comment":"Please report the random seeds used for the custom SportsPose and WorldPose splits, and specify the exact HRNet checkpoint used for detected keypoints, so that the protocol is reproducible.","section":"Section 4.1"},{"comment":"There are minor typos: 'eacy type' in Section 3.2 and 'Abalation' in the Table 5 caption.","section":"Typos"},{"comment":"The proposed model has 29.3 M parameters versus 18.9 M for MotionAGFormer-L; a parameter-matched baseline or a MotionAGFormer-L augmented with the same anatomy branches would help attribute the reported gain to the proposed components rather than to additional capacity.","section":"Overall comparison"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent incremental architecture paper with a clear internal ablation. My main reservation is not the architecture but the external validity of the SOTA claim: the custom splits and single-run results need to be addressed. If the held-out-game or official-split results confirm the trends, I would support acceptance after minor revisions; otherwise the paper should be reframed as a controlled comparison on a custom protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is an incremental but sensible architecture for 3D pose lifting on short sports clips: it takes the bone-direction/length features from Chen et al., adds a limb-fusion MLP that composites bones into biological and hyper-limbs, and runs cross-attention between bone and limb tokens inside a MotionAGFormer-style spatio-temporal transformer. That specific combination looks new, and the ablations show each piece contributes. They retrain all baselines on their own 27-frame protocol, so the internal comparison is controlled. The promise of code is a plus.\n\nThe soft spot is the evaluation protocol, and the stress-test note is right. The WorldPose split randomly assigns clips from each game to train and test, so clips from the same game—same camera setup, same players, possibly overlapping timestamps—appear on both sides. The model can exploit scene-specific cues. The reported 1.4 mm edge over MotionAGFormer-B on DET and 0.8 mm over D3DP on GT could plausibly shrink or reverse under a game-disjoint split. The SportsPose split is also custom (random 7 subjects, not shown to match official benchmarks), so the \"state-of-the-art\" label only applies to this internal setup. With single runs and no error bars, sub-millimeter to 1.5 mm differences are within typical run-to-run variance. I trust the ablations as evidence that bone/limb features help on this protocol, but not as evidence of benchmark SOTA. Two replication gaps: lambda_v in the loss is never specified, and the manual limb groupings are not listed.\n\nThis is a reasonable paper for a conference workshop or a main conference with revision. The core idea is not revolutionary, but it is cleanly executed and the ablations are honest. The main fix is to redo WorldPose with a held-out-game split, report variance across seeds, and either use the official SportsPose split or drop the SOTA wording. If the authors do that, the result becomes credible. I would send it to peer review, conditionally, and expect heavy revision on the evaluation. For a reading group, it is a decent example of anatomy-aware feature engineering, but not a must-read.","headline":"A clean incremental architecture for sports 3D HPE whose internal ablations are convincing, but the SOTA claim rests on leaky custom splits and single runs.","tokens_in":17210,"tokens_out":1850,"would_cite":false,"duration_ms":22034,"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":"KASportsFormer reports that explicit bone and limb tokens improve 3D pose estimation on short sports clips.","keywords":["3D human pose estimation","sports pose estimation","short video clips","bone and limb features","cross-attention transformer","kinematic anatomy","MPJPE","2D-to-3D lifting"],"falsifier":"Take KASportsFormer and MotionAGFormer-L, train and test both on the standard SportsPose and WorldPose evaluation splits with exactly 27-frame clips whose source subjects and game timestamps are fully disjoint from training, and compare MPJPE; if the difference falls within run-to-run noise or reverses, the short-video anatomy advantage is an artifact of the custom split rather than a property of the model.","tokens_in":16186,"feed_emoji":"🏃","tokens_out":8957,"duration_ms":99266,"temperature":0.7,"pith_summary":"This paper sets out to show that explicit kinematic anatomy improves 3D human pose estimation in short sports videos, where decisive motions such as a shot or a pass unfold over only a few frames and temporal context is scarce. It builds KASportsFormer, a transformer that lifts 2D keypoints to 3D using three parallel feature streams: raw joints, bone vectors, and fused limb tokens. Bone tokens query limb tokens through spatial and temporal cross-attention, while the joint stream keeps fine-grained locations, and the streams are blended adaptively at every layer. On 27-frame clips from SportsPose and WorldPose, the authors report state-of-the-art mean per-joint position error (MPJPE) of 58.0 mm and 34.3 mm, respectively, with the largest gains over MotionAGFormer-L under noisy HRNet-detected keypoints. If the result holds, anatomy-aware tokenization offers an architecture-level way to reduce the short-video weakness of current pose transformers.","feed_headline":"Bone and limb tokens cut 3D pose error on short sports clips","feed_subtitle":"It reports state-of-the-art MPJPE of 58.0 mm and 34.3 mm on short SportsPose and WorldPose clips.","key_machinery":"The load-bearing machinery is the anatomy extraction and fusion pair plus a cross-modal attention mixer. BoneExt converts a 2D pose into directed bone vectors by chaining joints from the pelvis, recording each bone's normalized direction and length plus an averaged ultra-bone; LimbFus then groups these bones into limbs, each encoded by a small MLP that treats x-direction, y-direction, and length separately, and supplements anatomically inspired limbs with non-biological hyper-limbs. The Anatomy Mixer runs multi-head cross-attention first over space then over time, with bone tokens as queries and limb tokens as keys and values, so the network reconciles single-bone kinematics with limb-level motion. A joint stream with self-attention and graph convolution preserves coordinate-level detail, and adaptive blending merges the three streams at every layer before the regression head.","core_discovery":"The paper's central claim is that the human skeleton carries two complementary kinematic descriptions—bones as directed vectors with length and direction, and limbs as fused groups of bones—and that forcing a pose transformer to exchange information between these descriptions makes it more accurate on brief sports actions. KASportsFormer decomposes each 2D frame into bone vectors, averages them into an ultra-bone, and then composes the bones into limb tokens, including non-anatomical hyper-limbs such as left shoulder to right hip. Inside each transformer layer, the bone stream supplies queries to a spatial then temporal multi-head cross-attention whose keys and values come from the limb stream, while a parallel joint stream runs self-attention and graph-convolution mixing for fine positions. The three outputs are adaptively blended and passed onward. The authors report that this design beats previous methods on both SportsPose and WorldPose at 27-frame input length, with the clearest visual improvements at limb junctions and fast-moving extremities.","pith_inferences":["Editorial extension: because the reported gains rest on the paper's own random subject split of SportsPose and random per-game split of WorldPose, a re-run under the datasets' standard evaluation protocols could shift the margins of 0.1 to 1.5 mm, and clip-level leakage should be checked before accepting the numbers as benchmark state of the art.","Editorial extension: if bone-length stability is the active ingredient, the same bone-plus-limb tokenization should transfer to other short-clip pose tasks such as dance, rehabilitation, or sports science workflows, and could be tested by swapping in a non-sports short-clip benchmark.","Editorial extension: the design keeps the limb stream fixed to the raw input rather than feeding it blended features; ablating this choice would reveal whether a frozen kinematic context is essential or whether letting limbs adapt would help further."],"forward_implications":["Short-clip robustness: the gains on 27-frame inputs imply that anatomy streams can partially substitute for the temporal context that longer clips would provide, so sports analysis systems can use shorter windows without losing as much accuracy.","Noise tolerance: larger improvements under HRNet-detected keypoints than under ground-truth projected keypoints suggest the anatomy representation is particularly useful when the 2D inputs are noisy, a common condition in broadcast sports video.","Action-specific behavior: the action breakdown shows the biggest MPJPE reductions on throw, tennis, jump, and volley, while soccer detection input slightly regresses; the authors attribute this to limited perception of limb acceleration in kicking.","Module simplicity: the ablation finding that a small 16-hidden-unit MLP performs best implies the limb fuser does not need a large network, so the anatomy branch adds modest parameters relative to the baselines.","Clean-input performance: on WorldPose ground-truth 2D keypoints, KASportsFormer reaches 8.5 mm MPJPE, indicating the benefit is not only a robustness effect under noise but also improves reconstruction given clean poses."],"supporting_citations":[{"why":"Supplies the SportsPose dataset, the short sports footage and 3D joint annotations on which the main comparison and ablations are run.","marker":"[10]"},{"why":"Supplies the WorldPose dataset, FIFA World Cup broadcast clips with SMPL annotations used as the second real-scene benchmark.","marker":"[13]"},{"why":"MotionAGFormer, the transformer-GCN baseline whose architecture, losses, and short-clip setup KASportsFormer extends and outperforms.","marker":"[21]"},{"why":"Anatomy-aware bone decomposition; introduces the bone direction and length pose representation that BoneExt instantiates.","marker":"[3]"},{"why":"HDFormer; contributes the idea of hyper-bone/joint and shortest-path connections that LimbFus adapts for non-biological limbs.","marker":"[2]"},{"why":"MotionBERT; supplies the adaptive token-blending strategy used to merge the joint, bone, and limb streams each layer.","marker":"[56]"},{"why":"HRNet; the 2D pose detector that produces the detected-keypoint inputs on both datasets.","marker":"[35]"},{"why":"KTPFormer; provides prior evidence that kinematic and trajectory priors improve 3D pose estimation, motivating the anatomy branch.","marker":"[29]"}],"fun_headline_variants":["Bone-limb fusion sharpens 3D pose on fleeting sports moves","KASportsFormer: bone tokens beat sports pose blur","Bones plus limbs cut pose error on fast sports clips","Anatomy-aware transformer improves short-sport 3D pose","Splitting bones from limbs boosts sports pose accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim stands on the authors' own random subject split of SportsPose and random per-game split of WorldPose; if these splits are unrepresentative of standard evaluation or leak subject, scene, or temporal context between training and test clips, the reported gains of roughly 0.1 to 1.5 mm may not reproduce outside this setup.","fun_headline_variants_meta":{"raw":{"variants":["Bone-limb fusion sharpens 3D pose on fleeting sports moves","KASportsFormer: bone tokens beat sports pose blur","Bones plus limbs cut pose error on fast sports clips","Anatomy-aware transformer improves short-sport 3D pose","Splitting bones from limbs boosts sports pose accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1431,"prompt_tokens":1010,"completion_tokens":421,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":626,"tokens_out":421,"duration_ms":5242,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:15:58.890824+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take KASportsFormer and MotionAGFormer-L, train and test both on the standard SportsPose and WorldPose evaluation splits with exactly 27-frame clips whose source subjects and game timestamps are fully disjoint from training, and compare MPJPE; if the difference falls within run-to-run noise or reverses, the short-video anatomy advantage is an artifact of the custom split rather than a property of the model.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SportsPose dataset, the short sports footage and 3D joint annotations on which the main comparison and ablations are run."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the WorldPose dataset, FIFA World Cup broadcast clips with SMPL annotations used as the second real-scene benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MotionAGFormer, the transformer-GCN baseline whose architecture, losses, and short-clip setup KASportsFormer extends and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Anatomy-aware bone decomposition; introduces the bone direction and length pose representation that BoneExt instantiates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HRNet; the 2D pose detector that produces the detected-keypoint inputs on both datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"KTPFormer; provides prior evidence that kinematic and trajectory priors improve 3D pose estimation, motivating the anatomy branch."}],"review_version":1}