{"id":"af9c882c-c202-4bca-920f-86909fb1e00c","arxiv_id":"2504.18332","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SSD-Poser reconstructs full-body SMPL poses from three sparse HMD signals with a hybrid state-space/attention encoder and frequency-aware decoder, reporting state-of-the-art accuracy at real-time speed on AMASS.","lead":"SSD-Poser predicts a full-body avatar pose from only the head and hand tracking signals that VR headsets provide. It combines a Mamba-2 state-space encoder with attention and a frequency-aware decoder to run in real time while reporting lower pose error than prior methods on AMASS.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AMASS random-split protocol and missing noise model leave the SOTA accuracy claim unverified; temporal leakage could explain the reported gains.","rationale":"The reader's weakest assumption identifies the same structural risk: the benchmark does not establish that the reported accuracy transfers to real HMD conditions. My stress-test sharpens this into a concrete, checkable flaw: the paper's Section 5.3 split description is ambiguous, and Section 4.1 omits the input-synthesis noise model. Since the accuracy margins over the strongest baseline are small (0.13-0.28 cm MPJPE), even mild temporal leakage or split differences could account for the reported advantage. This is more load-bearing than the paper's jitter results (where SSD-Poser is not best) or its inference-speed claim (Table 3 shows KCTD at 0.005 s vs 0.007 s), because those are secondary to the headline accuracy and can be fixed by rephrasing; the accuracy claim depends on a protocol that is currently unspecified. I therefore keep the reader's CONDITIONAL verdict: the paper is plausible but the empirical contribution cannot be fully assessed until the protocol is released and the check is run. No ad hominem is intended; the issue is missing methodological detail, not suspected misconduct.","tokens_in":11856,"tokens_out":7516,"duration_ms":70904,"concrete_test":"Obtain the authors' exact train/test sequence IDs and input-synthesis code. Retrain SSD-Poser and SAGE on a subject-exclusive split of the same 14 AMASS subsets, adding realistic Gaussian noise to the head/hand positions and rotations (e.g., 1-2 cm position noise and 1-2 degree orientation noise) and reporting mean plus or minus standard deviation over 5 random seeds. If the MPJPE gap versus SAGE drops below 0.1 cm or the standard deviations overlap, the SOTA accuracy claim is not robust to the evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SSD-Poser surpasses state-of-the-art methods in reconstruction accuracy rests entirely on Tables 1-2, which are produced from an underspecified AMASS evaluation protocol. Section 5.3 states only that the data are 'randomly divided into training and test sets with 90% and 10% of the data', without specifying whether the split is by sequence, subject, or frame, and Section 4.1 defines the sparse inputs (Eq. 6) but never describes how they are synthesized from AMASS or whether sensor noise is added. Because AMASS frames are temporally autocorrelated, a frame-level random split leaks near-duplicate poses into the training set; even a sequence-level split that mixes subjects can inflate accuracy through style leakage. The reported margins are small (MPJPE 2.67 vs 2.95 cm, MPJRE 2.18 vs 2.41 degrees against SAGE in Table 2), and no error bars, multiple seeds, or real-device evaluation are provided. The paper also does not state whether baseline numbers were re-computed under the same split or copied from prior papers. Until the split and input-synthesis protocol are pinned down, the SOTA accuracy claim is not falsifiable and may not transfer to noisy HMD tracking.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SSD-Poser, a lightweight architecture for full-body pose estimation from sparse HMD tracking signals (head and two hands). The method combines a hybrid encoder (State Space Attention Encoder, SSAE) that mixes a state-space-duality (SSD/Mamba-2) block with multi-head attention, and a Frequency-Aware Decoder (FAD) with parallel 1x1 and 1x5 convolutions to separate low- and high-frequency motion components. The model is trained end-to-end on AMASS with a loss combining rotation, position, and root orientation terms, and evaluated under two AMASS splits against prior HMD-based methods (AvatarPoser, DAP, AGRoL, KCTD, AvatarJLM, SAGE). The paper reports state-of-the-art accuracy (e.g., MPJPE 2.67 cm, MPJRE 2.18 degrees on 14 AMASS subsets) with 7.34M parameters and 0.007 s average inference per sequence, and provides ablations showing the contribution of PSSB, the SSM block, L2 loss, and the FAD module.","tokens_in":12144,"tokens_out":2850,"duration_ms":27230,"significance":"If the reported results are reproducible and transfer to real HMD tracking, the paper makes a useful contribution: it demonstrates that a hybrid SSD/attention encoder plus a frequency-aware decoder can achieve competitive accuracy at lower computational cost than transformer- and diffusion-based baselines, which is relevant for real-time AR/VR applications. The architecture is simple, lightweight, and the ablation study gives some support for the role of the FAD in improving smoothness (MPJVE and Jitter). However, the empirical verification is currently not strong enough to support the central SOTA claim: the evaluation protocol is underspecified in several load-bearing ways, and no uncertainty quantification is provided. The architectural novelty is incremental rather than fundamentally new, but the combination is reasonable.","major_comments":[{"comment":"The AMASS splitting protocol is not specified at the granularity needed to evaluate the results. The text says 'The datasets in both settings are randomly divided into training and test sets with 90% and 10% of the data,' but does not say whether the split is by frame, by sequence, or by subject. Because AMASS frames are temporally autocorrelated, a frame-level random split leaks near-duplicate poses into the training set and can substantially inflate accuracy; even a sequence-level split that mixes subjects can cause style leakage. Since the SOTA claim in Tables 1-2 rests entirely on this split, please specify the exact split unit, and preferably re-run the comparison with a subject/sequence-exclusive split and report results over multiple random seeds.","section":"Section 5.3 (Evaluation Results)"},{"comment":"The synthesis of sparse inputs from AMASS is not described. Equation (6) defines the input format (positions, 6D rotations, linear and angular velocities for head and hands), but the paper never states how these signals are derived from the SMPL parameters, whether the same synthesis is applied to all baselines, or whether any sensor noise or missing-data model is used. Without this, the numbers in Tables 1-2 are not reproducible, and the claimed suitability for 'real noisy HMD tracking' is unverified. Please add a precise description of the input-generation pipeline, including coordinate frames, sampling, and any noise, and specify whether baseline methods were re-run under the identical protocol or whether numbers were taken from prior papers.","section":"Section 4.1 (Problem Formulation) and Section 5.3"},{"comment":"The reported accuracy margins over the strongest baseline are small (e.g., MPJPE 2.67 vs 2.95 cm, MPJRE 2.18 vs 2.41 degrees against SAGE in Table 2), yet no error bars, multiple seeds, or significance tests are provided. Given the underspecified split and input synthesis, the claim in Section 6 that the model 'surpasses state-of-the-art approaches in reconstruction accuracy' is not currently falsifiable. Please provide confidence intervals across at least 3-5 seeds and, if possible, a per-subject breakdown to show that the improvement is consistent rather than driven by a particular split.","section":"Section 5.3 and Table 2"},{"comment":"The hyperparameters T=96 and the number of SSAE blocks (4) appear to be selected using the same test split that is used for the final reported numbers (Tables 1-2 and the ablation Tables 5-6). If no separate validation set was held out, the reported 'optimal' configuration may be overfit to the test set. Please clarify the model-selection procedure, or re-evaluate with a validation set, and report the final test numbers only for the configuration chosen on validation.","section":"Section 5.4 (Ablation Study)"},{"comment":"The text states that in the FAD ablation 'most metrics remain relatively stable,' but Table 4 shows that removing FAD degrades MPJVE from 19.32 to 27.80 and Jitter from 8.19 to 23.22, which are substantial changes. The qualitative description contradicts the quantitative result. Please correct this description and discuss the role of FAD in light of the actual numbers.","section":"Section 5.4 and Table 4"}],"minor_comments":[{"comment":"The entry 'Avatar]LM' appears to be a typo for 'AvatarJLM;' please fix.","section":"Table 3"},{"comment":"'Muti-Head Attention' should be 'Multi-Head Attention.' Please also check for other spelling/grammar issues, e.g., 'in the second line, the side-kick pose (in the third line), .' in Section 5.3.","section":"Section 4.3 (State Space Attention Encoder)"},{"comment":"The description of the learning rate schedule is incomplete: it says the learning rate 'decays to 3e-5 after 200000 iterations,' but does not specify the decay schedule (step, cosine, etc.). Please provide the full schedule.","section":"Section 5.1 (Dataset and Implementation Details)"},{"comment":"The ablation over the number of SSAE blocks reports that 5 blocks achieves the best MPJRE/MPJPE (2.37/3.11) but the paper selects 4 blocks citing 'precision, smoothness, and model efficiency.' Please justify this choice quantitatively, e.g., by reporting the inference time or parameter-vs-accuracy trade-off for each setting.","section":"Section 5.4 / Table 6"},{"comment":"Several dataset references are missing full citations (e.g., ACCAD, SFU, Eyes Japan, KIT, MoVi). Please provide complete bibliographic entries consistent with the style used for other references.","section":"References"},{"comment":"The caption 'Visualization results of different actions compared with other state-of-the-art methods' would benefit from naming the specific actions and the color-map definition for the red error regions, which are described only in the text.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is submitted to a conference (ICMR '25) and the core idea is reasonable, but the evaluation section needs substantial strengthening before the SOTA claim can be accepted. I am particularly concerned that the split granularity and input synthesis are not described, and that the baseline numbers' provenance is unclear; the authors should also address the lack of uncertainty quantification. There is some overlap in authorship between this work and two of the baselines (KCTD and DAP); I do not view this as improper, but it makes the reporting of baseline protocols especially important. The novelty is moderate, but the real-time efficiency result may be of interest to the venue if the accuracy claims can be made reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible engineering paper, first to bring Mamba-2's SSD into HMD-based full-body pose estimation, with a clean hybrid encoder and a frequency-aware decoder that demonstrably helps smoothness. The ablations are the best part: removing FAD hurts MPJVE and jitter a lot, removing the SSM block hurts everything, and the sequence-length and block-count sweeps are informative. The architecture write-up is detailed enough to reimplement. That is real value.\n\nThe problem is the evaluation protocol. Section 5.3 says only that the data were \"randomly divided into training and test sets with 90% and 10%\" with no split unit. AMASS is mocap sequences with strong temporal autocorrelation. If the split is at frame level, windows from the same motion sequence appear in both train and test, and a temporal model can memorize near-duplicates. Given the margins over SAGE are small (MPJPE 2.67 vs 2.95 cm, MPJRE 2.18 vs 2.41 degrees), this is not a nit; it determines whether the headline \"surpasses state-of-the-art\" means anything. The paper also never specifies how the sparse head/hand inputs are synthesized from AMASS or whether sensor noise is added, so the numbers may not transfer to real HMD tracking.\n\nAdd to that: no error bars or seeds, no statement about whether baseline numbers were recomputed under the same split, and the GT jitter values in Tables 1 and 2 are nonzero and mutually inconsistent. These are all fixable with text, code, or a few reruns, but right now the empirical claim is not falsifiable.\n\nWho is this for? A practitioner who wants a lightweight real-time baseline architecture will get something usable. A reader who wants to believe the SOTA numbers should wait for the protocol to be pinned down. I would send it to peer review, because the architecture and ablations deserve referee time, but the first round should demand the split details, input synthesis description, error bars, and code or data release.","headline":"A coherent Mamba-2 + attention pose estimator with useful ablations, but the AMASS split and input-synthesis protocol are underspecified enough that the SOTA accuracy claim is not yet verifiable.","tokens_in":12662,"tokens_out":2995,"would_cite":false,"duration_ms":32159,"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":"SSD-Poser reconstructs full-body avatar motion from only head-and-hand tracking signals, reporting state-of-the-art accuracy on AMASS with 7 ms inference and 7.34M parameters.","keywords":["human pose estimation","sparse observations","state space duality","head-mounted display tracking","frequency-aware decoder","full-body motion reconstruction","AMASS","real-time inference"],"falsifier":"Record a person wearing an HMD plus a synchronized optical motion-capture suit, feed only the head and hand tracks to SSD-Poser, and compare lower-body MPJPE against the 4.80 cm reported on the AMASS 14-subset split; if the real-device lower-body error exceeds roughly twice the benchmark error, the transferability claim collapses.","tokens_in":11677,"feed_emoji":"🧍","tokens_out":5878,"duration_ms":53170,"temperature":0.7,"pith_summary":"SSD-Poser aims to prove that full-body avatar pose can be reconstructed in real time from the three sparse signals a head-mounted display already provides—head and two hands—without the heavy cost of Transformers or diffusion models. The paper's central claim is that a hybrid encoder pairing the linear-time State Space Duality (SSD) recurrence with multi-head attention, followed by a frequency-aware decoder, recovers lower-body motion accurately enough to beat prior methods on the AMASS benchmark while running at 0.007 seconds per sequence. If correct, this matters for AR/VR because it offers a lightweight backbone that could track avatars from existing HMD hardware rather than requiring IMU suits or cameras. The reported numbers include MPJRE of 2.18, MPJPE of 2.67 cm, and MPJVE of 15.25 cm/s on a 14-subset AMASS split.","feed_headline":"Head-and-hand signals reconstruct full body in 7 ms","feed_subtitle":"A hybrid state-space/attention encoder cuts lower-body error to 4.8 cm on AMASS while keeping 7.34M parameters.","key_machinery":"The load-bearing object is the hybrid State Space Attention Encoder (SSAE), which pairs a Pose State Space Block (PSSB)—a gated SSD update that propagates latent state through time—with a standard multi-head attention module, so long-range temporal dynamics are handled at linear cost while attention supplies inter-joint context. The second enabling mechanism is the Frequency-Aware Decoder (FAD), whose Frequency-Aware Feature Extractor passes features through 1x1 and 1x5 convolutions to isolate low-frequency smooth motion from high-frequency detail, then concatenates and refines them before the final joint regression. Together they convert 54-dimensional sparse HMD observations into 132-dimensional SMPL pose parameters (22 joints, 6D rotations each).","core_discovery":"On its own terms, the paper establishes that a state-space/attention hybrid can outperform dedicated generative and Transformer pose estimators on sparse-input full-body tracking. SSD-Poser encodes sequences of 96 frames of head/hand positions, rotations, and velocities through four State Space Attention Encoders (SSAE), each combining a Pose State Space Block (PSSB) based on the SSD recurrence $h_t = A_t h_{t-1} + B_t x_t$, $y_t = C_t^T h_t$ with a multi-head attention module; a Frequency-Aware Decoder then separates low- and high-frequency motion features with 1x1 and 1x5 convolutions before regressing SMPL joint rotations. The paper reports that this architecture beats the compared methods on both a 3-subset and a 14-subset AMASS protocol, with the largest gains on lower-body and position errors, while using 7.34M parameters and an average inference time of 0.007 seconds per sequence. The velocity metrics (MPJVE of 15.25 cm/s and jitter of 6.73 on the larger split) support the smoothness claim.","pith_inferences":["The same SSAE hybrid recipe could be applied to other sparse-sensor pose tasks, such as 4-IMU lower-body tracking, where linear-time state propagation may help long sequences more than attention alone.","Treating the frequency split as a learnable inductive bias suggests a testable extension: replacing the fixed 1x5 temporal kernel with a learnable or multi-scale bank should show whether the gain comes from frequency separation per se or from the extra receptive field.","The random 90/10 AMASS split reports synthetic noise-free HMD observations; a cleaner test would be to train on AMASS and evaluate on real HMD recordings with optical ground truth, where distribution shift is likely to lower the reported accuracy."],"forward_implications":["Real-time avatar tracking becomes feasible on HMD hardware without external sensors: 0.007 seconds per 96-frame sequence is well below interactive frame budgets.","The accuracy gains concentrate in lower-body and position errors (Lower PE of 4.80 cm versus 5.37 cm for the best diffusion baseline on the 14-subset split), reducing the most visible artifact in sparse tracking.","The lightweight 7.34M-parameter footprint makes SSD-Poser a practical backbone for on-device or embedded AR/VR systems.","The frequency-aware decoder's effect on MPJVE and jitter suggests that separating low- and high-frequency motion is a useful general smoothing strategy for pose regression."],"supporting_citations":[{"why":"Supplies the State Space Duality framework and the SSM recurrence that the PSSB block adapts.","marker":"[10]"},{"why":"Defines the sparse HMD input representation (position, rotation, velocities) and the AvatarPoser baseline/protocol the paper follows.","marker":"[22]"},{"why":"Provides the AMASS motion-capture dataset used for training and both evaluation splits.","marker":"[31]"},{"why":"Defines the SMPL body model whose first 22 joints are the output representation.","marker":"[29]"},{"why":"SAGE is the strongest diffusion-based baseline in the 14-subset comparison and the source of the second evaluation protocol.","marker":"[14]"},{"why":"AGRoL is a diffusion baseline and shares the first evaluation protocol and smoothness-oriented metrics.","marker":"[13]"},{"why":"AvatarJLM is a Transformer-based joint-level baseline the paper compares against in both settings.","marker":"[49]"},{"why":"Provides the jitter metric used to quantify motion smoothness and a physics-aware baseline lineage.","marker":"[47]"}],"fun_headline_variants":["State-space hybrid tracks full body from head and hands","SSD-Poser: full-body pose from sparse signals in 7 ms","Fast avatar pose from head and hands with state-space","7 ms full-body pose from head and hand signals","State-space attention beats generative models on AMASS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire result depends on the premise that head and hand signals alone carry enough information to infer unconstrained lower-body motion, and that AMASS's synthetic sparse observations are a faithful stand-in for real HMD tracking noise and user behavior.","fun_headline_variants_meta":{"raw":{"variants":["State-space hybrid tracks full body from head and hands","SSD-Poser: full-body pose from sparse signals in 7 ms","Fast avatar pose from head and hands with state-space","7 ms full-body pose from head and hand signals","State-space attention beats generative models on AMASS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":2011,"prompt_tokens":985,"completion_tokens":1026,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":946}},"tokens_in":601,"tokens_out":1026,"duration_ms":9792,"temperature":1.0,"reasoning_tokens":946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:19:01.694827+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a person wearing an HMD plus a synchronized optical motion-capture suit, feed only the head and hand tracks to SSD-Poser, and compare lower-body MPJPE against the 4.80 cm reported on the AMASS 14-subset split; if the real-device lower-body error exceeds roughly twice the benchmark error, the transferability claim collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the sparse HMD input representation (position, rotation, velocities) and the AvatarPoser baseline/protocol the paper follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SAGE is the strongest diffusion-based baseline in the 14-subset comparison and the source of the second evaluation protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AGRoL is a diffusion baseline and shares the first evaluation protocol and smoothness-oriented metrics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AvatarJLM is a Transformer-based joint-level baseline the paper compares against in both settings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the jitter metric used to quantify motion smoothness and a physics-aware baseline lineage."}],"review_version":1}