{"id":"9483366f-8729-4ee8-8410-ca5e132b48e9","arxiv_id":"2504.14618","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"VM-BHINet combines a Vision Mamba block with an interaction feature module to recover two interacting hand meshes from one RGB image, reporting 5.44 mm MPVPE and 5.09 mm MPJPE on InterHand2.6M.","lead":"A new network for 3D hand mesh recovery from single images uses Mamba-style state space models to process two hands together, reporting lower error and lower compute than prior methods on the InterHand2.6M benchmark. The accuracy gain is modest and the core module is described only qualitatively, with no code or error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All claimed gains depend on baseline comparability; the HIC table shows a ~30% relative improvement over EANet that no public code can verify, and the abstract's '2-3%' claim contradicts the ~11% relative gains in Table II.","rationale":"I read the paper as a systems proposal: insert a VMBlock into an EANet-style interaction pipeline and report accuracy and efficiency gains. The strongest evidence is Table II and Table III, and the most load-bearing assumption is that the EANet rows are faithful, protocol-matched baselines. The reader identified exactly this concern, and I agree. The paper's internal inconsistency reinforces it: the abstract promises a 2-3% reduction while the tables show an order-of-magnitude larger relative gain, and the HIC margin over EANet is so large that protocol mismatch is a plausible explanation. No code, evaluation scripts, error bars, or formal specification of VMBlock are provided, so the results cannot be checked. This is not an ad hominem or a novelty dispute; the issue is that the central quantitative claim lacks independent support. A REJECT verdict remains appropriate, though the underlying idea could be salvageable if the authors release code and reproduce the baseline rows under identical splits and evaluation code.","tokens_in":14093,"tokens_out":3555,"duration_ms":33453,"concrete_test":"Download the released EANet weights and official evaluation code, evaluate on the HIC official test split using the paper's root-aligned MPVPE/MPJPE protocol, and compare against Table III's EANet row (41.34 / 45.01). If those numbers do not reproduce, the headline SOTA claim is unsupported. If they do reproduce, the remaining discrepancy is the abstract's 2-3% claim versus the roughly 11-12% relative gains in Table II; recompute those relative reductions from the reproduced numbers to determine which quantitative claim is accurate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that VM-BHINet surpasses state-of-the-art methods rests on the EANet comparison rows. Table III (Section IV-D) reports VM-BHINet MPVPE 28.98 versus EANet 41.34 on HIC, a roughly 30% relative improvement over a published method whose released model was supposedly evaluated with asterisks. Because VM-BHINet is an EANet-style pipeline whose IFEM is credited to [2], a gain this large is implausible unless the HIC baseline was evaluated under a different protocol (crop, root alignment, or split). Compounding this, the abstract claims a 2-3% error reduction, yet Table II shows MPVPE dropping from 6.16 to 5.44 and MPJPE from 5.73 to 5.09, approximately 11-12% relative reductions. No code or evaluation scripts are provided, so neither the baseline numbers nor VM-BHINet's own numbers can be independently verified. This is not a conceptual objection to using SSMs for hand reconstruction; it is a protocol-comparability and reproducibility problem in the headline evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VM-BHINet, a 3D interacting-hand mesh recovery network that introduces Mamba-style state space models into an EANet-like pipeline. The architecture consists of a ResNet-50 backbone, a Vision Mamba Interaction Feature Extraction Block (VM-IFEBlock) that combines Mamba blocks with the IFEM interaction module from the authors' prior work, a Hand Joint Feature Extractor, a Joint Vision Mamba Block, and a Dual Hand Parameter Regressor. The authors report ablation studies and comparisons on InterHand2.6M and HIC, claiming that VM-BHINet reduces MPJPE and MPVPE by 2-3% and significantly surpasses state-of-the-art methods while also reducing parameters and GFLOPs.","tokens_in":14356,"tokens_out":8075,"duration_ms":72202,"significance":"If the reported numbers are reliable, the efficiency gains are substantial: the full model uses 36.99M parameters and 12.97 GFLOPs compared with 136.22M parameters and 28.49 GFLOPs for the EANet baseline, with lower MPVPE and MPJPE. A positive feature is that the asterisked comparisons use the authors' released models rather than numbers copied from previous papers. However, the central SOTA claim is not yet substantiated because the headline percentage reduction is inconsistent with the tables, the HIC baseline rows contain internally impossible aggregate values, and the experimental protocol and code are not provided.","major_comments":[{"comment":"The abstract and the third contribution state that VM-BHINet reduces MPJPE and MPVPE by 2-3%, but Table II reports overall MPVPE dropping from 6.16 to 5.44, a relative reduction of about 11.7%, and overall MPJPE dropping from 5.73 to 5.09, about 11.2%. The same discrepancy appears in Table I. The 2-3% claim is therefore unsupported by the paper's own tables and should be corrected or explicitly justified.","section":"Abstract, Section I, Table II"},{"comment":"The HIC results for the asterisked baselines are internally inconsistent. For EANet, MPJPE is 43.97 for Single, 38.88 for Two, and 45.01 for All; for ACR, MPVPE is 46.02 for Single, 40.32 for Two, and 39.10 for All, and MPJPE is 45.38, 42.98, and 47.09 for Single, Two, and All respectively. If All is an aggregate of the Single and Two sets, these values are impossible because a weighted average cannot lie outside the range of its components. This makes the claimed HIC improvement (28.98 vs 41.34 MPVPE, 27.47 vs 45.01 MPJPE) unverifiable and points to a protocol or transcription problem in the baseline rows.","section":"Table III"},{"comment":"The HIC evaluation is not described. Section IV-A specifies training details only for InterHand2.6M, and Section IV-B merely says the model was evaluated on HIC. There is no information about HIC training epochs, learning rate schedule, batch size, input resolution, crop strategy, root alignment, or which split was used. Without these details, the large HIC gains cannot be assessed or reproduced. The authors should provide the full HIC protocol and, ideally, release code and evaluation scripts.","section":"Section IV-A, Section IV-B"},{"comment":"The no-module ablation row in Table I (136.22M parameters, 28.49 GFLOPs, MPVPE 5.61/6.87/6.16, MPJPE 5.28/6.17/5.73) is numerically identical to the EANet row in Table II, yet the paper never identifies this baseline as EANet. The authors must state explicitly what this baseline is, how it was trained, and confirm that VM-BHINet and all asterisked baselines were evaluated under identical data splits, augmentation, and evaluation code. Without this, the comparability assumption that supports the central SOTA claim is unverified.","section":"Table I vs Table II"},{"comment":"The core new component, VMBlock, is described only verbally. The paper does not provide the state-space equations, the selective scan configuration, the token ordering, or the manner in which the SSM is applied to 2D feature maps. As a result, the proposed block cannot be reimplemented from the text. The authors should give exact formulations or pseudocode for the VMBlock.","section":"Section III-B, VMBlock"}],"minor_comments":[{"comment":"There are typos in the title and in the Table I caption: 'VM-BHINet:Vision' is missing a space after the colon, and 'IFEB LOCK' should be 'IFEBlock'.","section":"Title and Table I caption"},{"comment":"The paragraph following Eq. (11) begins 'Whereθgt' without a space; this should be corrected to 'Where θgt'.","section":"Equation (11)"},{"comment":"The loss weights λ1 through λ9 are introduced but their numerical values are never reported. Since the loss weights directly affect the reported errors, the authors should list the values used.","section":"Equation (11)"},{"comment":"The paper should clarify whether all metrics come from a single training run and, if so, report variance across seeds or at least soften the word 'significantly' in the abstract.","section":"Section IV-D"},{"comment":"Because IFEM is explicitly credited to the authors' prior work [2], the paper should clearly delineate which parts of VM-IFEBlock are new and should include a comparison with InterHandNet [2] to justify the new contribution.","section":"Section III-B, IFEM"}],"recommendation":"major_revision","confidential_remarks":"The reader's reject verdict is defensible given the unverifiable baseline comparability and the impossible HIC aggregate values. I recommend major revision rather than outright rejection because the issues are addressable: correct the headline percentage, fix or re-evaluate the HIC table, supply the full training/evaluation protocol, clarify the baseline construction, and release code. If the authors cannot provide corrected baseline numbers or code, the paper should be rejected in a subsequent round."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a first-to-apply: swapping self-attention for a Mamba block in an EANet-style two-hand mesh recovery pipeline. That is a legitimate but narrow contribution. The efficiency gains are real and substantial; the accuracy story is where I'd push back.\n\nWhat the paper does well: the architecture is coherent and the ablation is consistent. On InterHand2.6M, adding the VM-IFEBlock and JVMBlock improves both MPVPE and MPJPE at every split, and the parameter count drops from 136M to 37M with FLOPs roughly halved. That is a concrete efficiency win. The authors also openly credit IFEM to their own prior work, which I appreciate—they are not hiding the provenance of the main interaction module.\n\nNow the soft spots, in order of severity. First, the HIC numbers are not credible as reported. VM-BHINet goes from 41.34 to 28.98 MPVPE, a ~30% relative improvement over the EANet baseline, while the InterHand gain is only ~11%. That gap smells like a protocol mismatch—crop size, root alignment, evaluation script, or a different test split. The asterisk says the baseline came from the authors' released model, but without code or evaluation scripts I cannot verify that both models were treated identically. Second, the abstract says 2-3% error reduction, but Table II shows MPVPE dropping from 6.16 to 5.44 and MPJPE from 5.73 to 5.09, roughly 11-12% relative. That mismatch is not a small typo; it is the headline claim, and it should have been caught. Third, the core VMBlock is never mathematically specified. The text says it combines SSM, Conv, MLP, LayerNorm, and Linear, and there is a figure, but no equations, no sequence order, no scan direction, no state dimension. That makes the paper's central contribution impossible to reproduce from the write-up alone.\n\nThe stress-test concern holds up: the central SOTA claim does rest on baseline comparability, and the HIC table is the weakest link. The absence of error bars is minor by comparison—these metrics are usually deterministic after training, so I would not overweight that.\n\nWho is this for? Someone working on efficient hand mesh recovery might find the architecture worth trying, especially the Mamba-for-joints idea. But the evaluation needs to be cleaned up before the accuracy claim is believable.\n\nRecommendation: I would not accept this as-is, but I would send it to peer review rather than desk-reject. The efficiency result is concrete, the first-to-apply point is publishable if the numbers are reproducible, and a serious referee could request code, a corrected abstract, and a proper HIC baseline check. With that, it could become a decent workshop or short-conference paper.","headline":"First-to-apply Mamba for two-hand mesh recovery; efficiency gains look real, but the headline accuracy numbers rest on an implausible HIC comparison and a sloppy abstract, so the SOTA claim needs verification before I'd trust it.","tokens_in":14918,"tokens_out":2124,"would_cite":false,"duration_ms":21386,"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":"VM-BHINet claims that replacing attention with Mamba-style state-space blocks in 3D interacting-hand mesh recovery lowers MPVPE and MPJPE by 2–3% while using a fraction of the compute.","keywords":["3D hand mesh recovery","bimanual hand interaction","state space model","Mamba","Vision Mamba","InterHand2.6M","MANO","monocular RGB"],"falsifier":"Run EANet's released model and VM-BHINet on the official InterHand2.6M test split using identical training data, augmentation, and evaluation code; if EANet's reproduced MPVPE/MPJPE differ from the 6.16/5.73 mm reported in Table II, or if the gap to VM-BHINet's 5.44/5.09 mm shrinks below the claimed 2–3%, the central SOTA claim is not established.","tokens_in":13874,"feed_emoji":"🖐️","tokens_out":8727,"duration_ms":67682,"temperature":0.7,"pith_summary":"The paper aims to establish that state space models—specifically the Mamba variant—can replace much of the self-attention machinery in monocular 3D interacting-hand reconstruction, improving accuracy while cutting parameters and FLOPs. It proposes VM-BHINet, whose core Vision Mamba Interaction Feature Extraction Block (VM-IFEBlock) couples a Mamba-based Vision Mamba Block with a non-local interaction module between left- and right-hand features. On InterHand2.6M the authors report overall MPVPE of 5.44 mm and MPJPE of 5.09 mm, roughly 2–3% better than the EANet baseline (6.16 and 5.73 mm), with 36.99M parameters and 12.97 GFLOPS versus EANet's 136.22M and 28.49. On the HIC dataset they report a much larger gap (overall MPVPE 28.98 mm vs. 41.34 mm). If the comparison holds, this is a concrete step toward faster, cheaper bimanual hand tracking for AR/VR and human-computer interaction.","feed_headline":"State-space model shrinks two-hand 3D reconstruction error by 2-3%","feed_subtitle":"VM-BHINet reports 5.44 mm MPVPE and 5.09 mm MPJPE on InterHand2.6M at under a third of EANet's compute.","key_machinery":"The load-bearing component is the Vision Mamba Interaction Feature Extraction Block (VM-IFEBlock), which stacks a Vision Mamba Block (VMBlock) and an Interaction Feature Extraction Module (IFEM). VMBlock uses a state space model—a sequence model that propagates a hidden state across positions at linear cost—alongside convolutions, MLPs, LayerNorm, and linear layers to capture global dependencies; IFEM then applies non-local attention between left- and right-hand features to encode their interaction. A second application of VMBlock to joint features (Joint Vision Mamba Block, JVMBlock) refines per-hand joint representations before the Dual Hand Parameter Regressor maps them to MANO pose and shape parameters and relative translation, trained with a weighted sum of L1 losses over pose, shape, joints, meshes, and translation.","core_discovery":"VM-BHINet claims to be the first framework to bring state space models into 3D interacting hand mesh recovery. The network takes a single RGB image, extracts left/right features with a ResNet-50 backbone, concatenates them, and processes them through VM-IFEBlock: a VMBlock built on the Mamba selective SSM plus local convolutions, followed by an Interaction Feature Extraction Module that uses non-local attention to relate the two hands. The Hand Joint Feature Extractor and Joint Vision Mamba Block then refine per-hand joint features, and the Dual Hand Parameter Regressor predicts MANO pose and shape parameters plus the relative 3D translation between hands. The paper reports that this architecture lowers both per-vertex and per-joint errors on InterHand2.6M to 5.44 and 5.09 mm overall, surpassing EANet's 6.16 and 5.73 mm while using 27% of its parameters and 46% of its FLOPs; on HIC it reports even larger absolute gains. The central claim is that replacing quadratic self-attention with linear-time SSM-based blocks is enough to improve both accuracy and efficiency in tightly interacting hands.","pith_inferences":["The 2–3% margin on InterHand2.6M is small enough that protocol differences (data split, augmentation, root alignment, evaluation code) could flip it; a re-run of EANet's released model under exactly VM-BHINet's training protocol would settle whether the SSM blocks are the cause of the gain.","Table I's odd parameter pattern—both blocks together giving 36.99M params while either block alone gives 61.15M or 112.07M—suggests the architecture also changes the backbone or parameter sharing; readers should verify what the 'Contrast' row actually is before attributing the saving to SSMs alone.","The paper's own Limitation & Future Work section states that robustness under complex backgrounds/extreme lighting, real-time speed, cross-domain generalization, and multi-hand scenarios remain open; the reported numbers are therefore a benchmark result, not a deployment claim.","A natural next experiment would be to drop VM-IFEBlock into an existing attention-based hand reconstruction model as a plug-in block and measure error change, isolating the SSM effect from the rest of the architecture."],"forward_implications":["If the reported numbers hold, SSM-based blocks can replace attention blocks in interacting-hand reconstruction without sacrificing accuracy—VM-BHINet's overall MPVPE/MPJPE on InterHand2.6M are 5.44/5.09 mm vs. EANet's 6.16/5.73 mm.","The parameter and FLOP reductions (36.99M/12.97 vs. 136.22M/28.49 for EANet) make real-time or embedded bimanual tracking a more plausible target for AR/VR and HCI.","The large HIC gains (overall MPVPE 28.98 mm vs. EANet's 41.34 mm) suggest, if reproducible, that the interaction feature extraction transfers beyond the InterHand2.6M capture setting.","Combining VM-IFEBlock and JVMBlock is reported to be necessary: ablation with only one of the two raises error and parameters, so the interaction between the two blocks is part of the claimed mechanism."],"supporting_citations":[{"why":"Provides the InterHand2.6M dataset and the InterNet baseline used for training and evaluation.","marker":"[34]"},{"why":"The NeuralAnnot annotator, cited alongside [34] when describing InterHand2.6M's ground-truth generation.","marker":"[33]"},{"why":"Mamba, the selective state-space model that VMBlock builds on for linear-time global feature modeling.","marker":"[11]"},{"why":"EANet, the state-of-the-art method whose released model supplies the main baseline numbers in Table II and Table III.","marker":"[38]"},{"why":"The InterHandNet interaction feature extraction design that IFEM adapts.","marker":"[2]"},{"why":"MANO, the parametric hand model used to turn predicted pose and shape parameters into output meshes.","marker":"[40]"},{"why":"ResNet-50 backbone used for initial image feature extraction.","marker":"[16]"},{"why":"The HIC dataset used for the cross-dataset evaluation in Table III.","marker":"[46]"}],"fun_headline_variants":["First SSM for two-hand 3D mesh recovery, 2-3% error cut","State-space model for hand mesh: 2-3% better, 46% less compute","VM-BHINet uses SSM to beat EANet in hand mesh accuracy","Linear-time SSM gives 2-3% tighter hand meshes","SSM-based hand mesh recovery outperforms with fewer parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 'Contrast' baseline row in the ablation table is a faithful, comparably trained reproduction of EANet on the same InterHand2.6M split and evaluation; the paper does not release code or evaluation scripts, so that comparability is unverified.","fun_headline_variants_meta":{"raw":{"variants":["First SSM for two-hand 3D mesh recovery, 2-3% error cut","State-space model for hand mesh: 2-3% better, 46% less compute","VM-BHINet uses SSM to beat EANet in hand mesh accuracy","Linear-time SSM gives 2-3% tighter hand meshes","SSM-based hand mesh recovery outperforms with fewer parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000937,"raw_usage":{"total_tokens":4015,"prompt_tokens":963,"completion_tokens":3052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":2946}},"tokens_in":579,"tokens_out":3052,"duration_ms":22180,"temperature":1.0,"reasoning_tokens":2946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:44:07.323462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EANet's released model and VM-BHINet on the official InterHand2.6M test split using identical training data, augmentation, and evaluation code; if EANet's reproduced MPVPE/MPJPE differ from the 6.16/5.73 mm reported in Table II, or if the gap to VM-BHINet's 5.44/5.09 mm shrinks below the claimed 2–3%, the central SOTA claim is not established.","supporting_citations":[{"cited_title":"Interhand2","cited_arxiv_id":null,"evidence_quote":"Provides the InterHand2.6M dataset and the InterNet baseline used for training and evaluation."},{"cited_title":"Neuralannot: Neural annotator for 3d human mesh training sets","cited_arxiv_id":null,"evidence_quote":"The NeuralAnnot annotator, cited alongside [34] when describing InterHand2.6M's ground-truth generation."},{"cited_title":"Extract-and-adaptation network for 3d interacting hand mesh recovery","cited_arxiv_id":null,"evidence_quote":"EANet, the state-of-the-art method whose released model supplies the main baseline numbers in Table II and Table III."},{"cited_title":"Interhandnet: Enhancing 3d interacting hand mesh recovery through inter-hand feature extraction","cited_arxiv_id":null,"evidence_quote":"The InterHandNet interaction feature extraction design that IFEM adapts."}],"review_version":1}