{"id":"86feb1ea-6e44-45f1-99e8-03df9db97eac","arxiv_id":"2508.01850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ChairPose estimates full-body 3D seated pose from pressure maps and chair geometry, reaching 89.4 mm MPJPE on unseen user-plus-chair combinations.","lead":"ChairPose reconstructs a full 3D seated body pose from a pressure-sensing mat placed on a chair, without cameras or wearables, and works for chairs and people it has not seen in training. A smart generalist might care because passive pressure sensing could enable private, effortless posture monitoring for ergonomics, rehabilitation, and adaptive interfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 89.4 mm LOCUOCV figure is computed against SMPLer-X pseudo-ground-truth from a single 45° iPhone camera (§4.2–4.3); no marker-based or multi-view validation is reported, so systematic seated-pose bias in SMPLer-X would directly shift every reported accuracy number.","rationale":"The reader's weakest-assumption analysis is exactly right and I agree. The pseudo-GT issue is the most load-bearing because every quantitative claim in the paper is expressed relative to it; the sim-to-real transfer of PresSim and the conflicting augmentation percentages (42.8% vs 85.5%) are also concerns, but they affect the contribution's explanation rather than the measured quantity itself. The paper deserves credit for real-data evaluation, leave-one-out protocols, and an honest limitations section (arm/head, deformable surfaces, autoregressive drift). However, the absence of marker-based validation means the 89.4 mm number could be an artifact of SMPLer-X bias, and the inclusion of SMPLer-X as a 'vision baseline' at 11 mm is circular. This does not warrant rejection; it warrants a condition: provide marker-based or multi-view validation for at least a subset, or clearly reinterpret the metric as 'agreement with SMPLer-X' rather than absolute pose accuracy. Hence the reader's CONDITIONAL verdict stands; my read does not change it.","tokens_in":22021,"tokens_out":7252,"duration_ms":90414,"concrete_test":"Collect marker-based optical motion capture (Vicon/OptiTrack) simultaneously with the Sensing.Tex mat and the iPhone capture for at least one user on each of the four chairs, covering the 12 actions; run SMPLer-X on the same videos; (a) compute SMPLer-X-vs-marker MPJPE; (b) retrain/evaluate ChairPose using marker-based GT on that subset and recompute MPJPE. If the SMPLer-X-vs-marker error is comparable to or larger than 89.4 mm, or if ChairPose's marker-GT MPJPE differs by more than ~20 mm, the central claim is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All reported MPJPE/PA-MPJPE/MPVE numbers (Tables 5–7) are distances to SMPLer-X outputs, not to an independent ground truth. Section 4.2 fixes one iPhone TrueDepth camera at 45° front-left; §4.3 processes the video with SMPLer-X and cites the authors' own DMCB benchmark for accuracy, but that benchmark concerned loose garments and is not seated-pose validation. Seated configurations create chair-induced occlusions (lower legs, feet, parts of the thighs) that are exactly where monocular SMPLer-X is likely to have systematic bias. The paper also lists SMPLer-X as a comparison method at ~11 mm error, making the 'ground truth' and a compared system the same model. Finally, PA-MPJPE in LOCUOCV (88.7 mm) is nearly identical to MPJPE (89.4 mm), which is unusual and consistent with a shared pseudo-GT bias rather than genuine hip/global error. If SMPLer-X's seated-pose bias is on the order of the claimed 89.4 mm, the headline generalization result is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ChairPose, a pressure-mat-based system for full-body 3D seated pose estimation. It uses a two-stage generative pipeline: a VQ-VAE motion quantizer that discretizes SMPL pose sequences, and an autoregressive token predictor conditioned on pressure maps, a 3D chair point cloud, and previous pose tokens. Training is augmented by a physics-based simulation pipeline (PresSim plus ragdoll dynamics) that generates synthetic pressure-pose pairs for 12 additional chair models. The system is evaluated on a new dataset (TDSD) of 8 participants, 4 chairs, and 12 seated activities, under leave-one-user-out (LOUOCV), leave-one-chair-out (LOCOCV), and leave-one-(user+chair)-out (LOCUOCV) protocols. Reported MPJPE is 53.5 mm (LOUOCV), 88.2 mm (LOCOCV), and 89.4 mm (LOCUOCV), with large improvements over the non-quantized baseline and over re-implemented pressure-based baselines. Additional experiments cover ablations of the motion quantizer, data augmentation, volumetric center-of-mass estimation, posture monitoring, and activity recognition.","tokens_in":22259,"tokens_out":4439,"duration_ms":49185,"significance":"If the accuracy numbers are taken at face value, ChairPose is a substantial step for privacy-preserving seated pose estimation: it removes wearables and cameras, incorporates chair geometry explicitly, and demonstrates generalization to unseen users and chairs through physics-based simulated augmentation. The paper ships publicly available code and data, a sensible cross-validation protocol, and a detailed ablation study. The reported gains from motion quantization and from simulation-assisted augmentation are large and internally consistent across the three cross-validation settings. However, the central generalization claim rests on the validity of the pose ground truth, which is currently a monocular vision model output, and the paper overstates the sensor-side independence by requiring a 3D chair scan as input.","major_comments":[{"comment":"The ground-truth poses are produced by SMPLer-X from a single iPhone TrueDepth camera fixed at a 45-degree front-left viewpoint, and no independent validation is reported for seated configurations. Chair-induced occlusions of the lower legs, feet, and parts of the thighs are precisely where monocular SMPLer-X is most likely to have systematic bias, and any such bias directly shifts every reported MPJPE/PA-MPJPE/MPVE value. The cited DMCB benchmark (refs. [43,44]) concerned loose garments and does not establish SMPLer-X accuracy for seated poses with chair occlusion. Please add an independent validation subset, such as optical marker-based motion capture, multi-view triangulation, or a manually annotated joint sample, and report per-joint errors so the pseudo-ground-truth bias can be assessed.","section":"§4.3, Tables 5–7"},{"comment":"SMPLer-X is used both as the source of ground-truth annotations (Section 4.3) and as a comparison method in the evaluation tables, where it reports about 11 mm error. This makes the ~11 mm row a self-consistency measure rather than an independent accuracy benchmark, and it conflates the reference with a compared system. Either remove SMPLer-X from the comparison tables or explicitly relabel its numbers as pseudo-ground-truth agreement, not as an external reference.","section":"§5.1.2, Tables 5–7"},{"comment":"In LOCUOCV, PA-MPJPE (88.7 mm) is nearly identical to MPJPE (89.4 mm), whereas in LOUOCV PA-MPJPE is about 16 mm lower than MPJPE (37.9 vs. 53.5 mm) and in LOCOCV it is about 11 mm lower (77.1 vs. 88.2 mm). Such near-equality after Procrustes alignment is unusual and suggests a non-rigid per-joint bias that is not captured by a global similarity transform, or an artifact of the shared pseudo-ground-truth. Please report per-joint errors, the fitted Procrustes parameters, or a scatter analysis for this condition to explain why alignment does not reduce the error.","section":"Table 7"},{"comment":"The paper claims that the system operates 'solely on pressure sensing' and 'independently of chair geometry,' yet the P2P module takes as input the chair point cloud M ∈ R^{5000×3} (Section 3.2) and Section 4.1 describes acquiring 3D chair scans with a Ferret Pro scanner. This is an overclaim: the system is chair-geometry-aware and requires a scan at inference time, even though it does not require per-chair retraining. Please revise the abstract, the contribution list, and Table 1 to distinguish 'no per-chair retraining' from 'no chair-geometry input.'","section":"Abstract, §3.2, §4.1"},{"comment":"The simulation-assisted augmentation is a central contribution, but the paper does not report any direct measure of sim-to-real transfer, such as pressure-map distribution similarity or a real-synthetic domain gap analysis. The reported accuracy gains are consistent across cross-validation protocols, which is encouraging, but an explicit validation that synthetic pressure maps resemble real sensor readings on the four physical chairs would strengthen the claim that the physics-based pipeline is responsible for the improvement rather than merely adding data diversity.","section":"§5.1.2 (data augmentation ablation)"}],"minor_comments":[{"comment":"The metric notation in Eq. (10) is internally inconsistent: 'X is \\hat{p}_j' is not a proper definition, and the norm subscript is missing. Please rewrite the equation with consistent notation for prediction and ground truth.","section":"§5.1.1, Eq. (10)"},{"comment":"The codebook size is given as 1028, which is not a power of two and is an unusual choice; please confirm whether this is a typo for 1024 and, if not, explain the rationale.","section":"§4.5, Table 3"},{"comment":"The text states that the MQ+sequence mode has a fixed ~2 second latency 'regardless of input length,' but the source of this fixed delay is not explained given the 1-second quantization window; please clarify whether the delay comes from the autoregressive decoding, the sequence loss, or the temporal receptive field.","section":"§7.3, Figure 9"},{"comment":"Error accumulation is acknowledged as a limitation of the autoregressive decoder, but no quantitative drift analysis is provided; an error-over-time plot would make the severity of this issue concrete.","section":"§7.2.3"},{"comment":"The column 'Cross-surface generalization' is not defined in the table or the text; please clarify whether it indicates operation across chair shapes, planar surfaces only, or something else.","section":"Table 1"},{"comment":"The posture-monitoring experiment reports mean absolute errors of 4.2° for lumbar flexion and 3.8° for thoracic tilt, but the paper does not specify how the ground-truth angles were computed from the SMPL poses or how the predicted angles are derived; please add this detail for reproducibility.","section":"§6.2"}],"recommendation":"major_revision","confidential_remarks":"The heavy reliance on the same group's PresSim simulator as the synthetic training-data generator is a potential circularity risk; an independent simulator or a dedicated sim-to-real validation would make the data-augmentation claim more convincing. The term 'first' in the abstract and the comparison table should be checked against recent tactile-carpet and in-bed pressure methods to avoid overclaiming novelty. The paper otherwise has a solid evaluation structure and public code/data, which are assets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ChairPose is a genuine step forward for pressure-based seated pose estimation: the combination of VQ-VAE motion tokenization, PointNet chair-geometry conditioning, and PresSim-plus-ragdoll augmentation is new, and the three cross-validation protocols (LOUOCV, LOCOCV, LOCUOCV) are more thorough than this area usually gets. The consistent performance gains from quantization and physics-based augmentation—53.5 vs 195.2 mm MPJPE on LOUOCV, for example—make a credible case that the architecture works on held-out real data.\n\nThe soft spot is the ground truth. Section 4.2 describes one iPhone TrueDepth camera fixed at a 45° front-left angle; Section 4.3 says the raw video was processed with SMPLer-X to produce the 3D pose labels. There is no marker-based or multi-view validation anywhere. The paper invokes the authors' own DMCB benchmark to justify monocular accuracy, but that benchmark addressed loose garments, not seated postures with chair-induced occlusion. SMPLer-X then appears in Tables 5–7 as a comparison method at ~11 mm error, which is circular because it is scoring against its own outputs. And the PA-MPJPE of 88.7 mm compared with MPJPE of 89.4 mm in LOCUOCV is a red flag; Procrustes alignment usually removes substantial global error, and the near-identical values suggest a shared systematic bias in the pseudo-labels rather than clean, independent error.\n\nThere are smaller issues worth naming. The augmentation ablation is internally inconsistent: the naive augmentation's penalty is reported as 42.8% MPJPE increase in one passage and 85.5% in another. The code/data link is just 'available on Kaggle' with no handle, commit hash, or DOI, so the reproducibility promise cannot be checked from the arXiv PDF. PresSim is the authors' own simulator used to generate training data, which is fine, but the sim-to-real transfer to the actual TPE mat is never validated, so the simulation contribution rests on an unverified assumption.\n\nThe core method is likely sound, and the TDSD dataset is a useful new resource. The absolute error claims, however, should be treated as provisional until the pseudo-GT is independently validated on at least a subset of frames. I would send this to peer review with a request for multi-view or marker-based spot-checking, per-joint error reporting, and a fix for the ablation inconsistency. It is the right kind of paper for the UIST audience, and the flaws are addressable.","headline":"Novel and well-evaluated pressure-to-pose system whose absolute accuracy is undercut by monocular pseudo-ground-truth; deserves refereeing but needs independent validation.","tokens_in":22849,"tokens_out":4364,"would_cite":true,"duration_ms":46702,"reading_group":"yes","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 thin pressure mattress plus a one-time 3D scan of the chair can reconstruct full-body seated pose on users and chairs never seen during training, with a mean per-joint position error of 89.4 mm.","keywords":["seated pose estimation","pressure sensing","chair morphology","motion quantization","VQ-VAE","physics-based simulation","data augmentation","privacy-preserving sensing"],"falsifier":"Run a marker-based or multi-view motion-capture system on the same 12 activities across the four chairs, and compare ChairPose's predicted joints to those marker positions for held-out users and chairs; if the mean per-joint error against true marker-based ground truth is substantially above 89.4 mm, especially for arm and head joints, the central claim fails. A second direct test is to place the mattress on a soft, deformable seat and see whether the claimed chair-agnostic accuracy collapses.","tokens_in":21805,"feed_emoji":"🪑","tokens_out":9085,"duration_ms":90482,"temperature":0.7,"pith_summary":"ChairPose sets out to prove that full-body 3D seated pose can be reconstructed passively: a thin pressure mattress placed on a chair, together with a one-time scan of that chair, replaces cameras and wearable sensors. The paper claims this works on chairs and people the model has never seen, reporting a mean per-joint position error of 89.4 mm when both are held out. The system's two-stage design first quantizes body motion into discrete tokens, then predicts those tokens autoregressively from pressure dynamics and chair geometry, and a physics-based simulation pipeline supplies extra training pairs across 12 synthetic chair shapes. If the claim holds, continuous posture monitoring, center-of-mass estimation, and seated activity recognition become possible in privacy-sensitive settings with no cameras or body-worn sensors.","feed_headline":"Pressure mat reads full-body sitting pose on unseen chairs","feed_subtitle":"With only a thin mattress and one chair scan, it reports 89 mm joint error on chairs and users it has never seen, no cameras or wearables.","key_machinery":"The load-bearing mechanism is the two-stage generative pipeline. MotionQuantizer is a VQ-VAE with a U-Net backbone that compresses n-second chunks of SMPL joint positions and motion derivatives into codebook tokens, using exponential moving average updates and quantization dropout to keep the codebook stable. Pressure2Pose is an autoregressive classifier over those tokens, conditioned on an 80x28 pressure-map sequence, a 5000-point chair point cloud encoded by PointNet, and the previously predicted token, with a learned start token to begin generation; outputs are decoded back into continuous SMPL joint positions. The second mechanism is the simulation-assisted augmentation pipeline, which replaces a flat-floor pressure simulator with 12 chair CAD models and runs ragdoll physics on motion-capture sequences so the synthetic pressure-pose pairs obey plausible body-chair contact.","core_discovery":"The core claim is that pressure maps alone are a sufficient signal for continuous full-body 3D seated pose estimation, provided the model is told the chair's 3D shape and trained with physically grounded synthetic examples. The paper demonstrates this with a two-stage generative architecture whose MotionQuantizer turns SMPL pose sequences into a discrete codebook, and whose Pressure2Pose module classifies each next codebook token from pressure frames, a PointNet-encoded chair point cloud, and the previously predicted token. The reported results are 53.5 mm MPJPE for unseen users, 88.2 mm for unseen chairs, and 89.4 mm for unseen users and chairs together, down from 260.7 mm for a direct-regression baseline under the hardest protocol. The same reconstructed mesh yields clinically oriented outputs: volumetric center-of-mass estimation with lower error than direct regression, spine-angle monitoring with mean errors under 4.2 degrees, and activity recognition improved by fusing the predicted pose with raw pressure.","pith_inferences":["Extrapolating from the paper's four chair types, the strongest untested boundary is deformable seating: soft cushions and dynamic surfaces change pressure patterns, so putting the mattress on a plush sofa would be the natural next experiment.","Because the ground truth comes from monocular vision extraction rather than marker-based capture, the 89.4 mm figure measures agreement with that vision model; a multi-view or marker-based validation would separate sensor ambiguity from pseudo-ground-truth bias.","The synthetic data pipeline is a reusable asset: the same ragdoll-plus-pressure simulation could generate training pairs for other chair-embedded tasks such as fall detection, occupancy sensing, or driver monitoring without new real data collection.","The tokenization approach suggests a path to cross-modal supervision: if pressure-to-token classification works, the same codebook could accept pose context from text or audio, enabling pose generation from multiple sparse modalities."],"forward_implications":["If the reported numbers hold, a single flexible mattress plus a 3D chair scan can replace cameras and IMUs for seated pose monitoring, working on chairs and users the model was never trained on.","The full-body SMPL mesh it outputs supports downstream tasks directly: center-of-mass estimation, spine angle monitoring, and seated action recognition all improve over pressure-only or direct-regression baselines.","The fixed roughly two-second latency of the accurate mode suits offline clinical or ergonomic assessment, while the fast baseline mode serves interactive, latency-critical uses at lower accuracy.","Physics-based synthetic augmentation, not simple image-style augmentation, is what drives generalization: the paper reports error reductions up to 46 percent from adding simulated pressure-pose pairs to the training set."],"supporting_citations":[{"why":"supplies the monocular SMPL extraction model used to label the real dataset with pseudo ground truth.","marker":"[6]"},{"why":"provides the physics-based pressure simulation pipeline that ChairPose extends from flat floors to non-planar chair surfaces.","marker":"[42]"},{"why":"supplies the ragdoll dynamics used to make synthetic characters interact plausibly with each chair geometry.","marker":"[37]"},{"why":"provides the VQ-VAE motion-quantization architecture that MotionQuantizer adapts from the TMD line of work.","marker":"[40, 41]"},{"why":"supplies the benchmark whose controlled-condition results justify using monocular vision-based pose as ground truth.","marker":"[43]"},{"why":"defines the SMPL body model whose pose parameters and mesh the pipeline quantizes, regresses, and evaluates.","marker":"[28]"}],"fun_headline_variants":["ChairPose: full-body pose from pressure alone, unseen chairs","Pressure maps plus chair shape estimate seated pose","Simulation-trained pressure system tracks sitting pose","No wearables: pressure mat reads pose on novel chairs","89 mm pose error from pressure on unseen users and chairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy numbers assume the monocular vision reconstructions used as ground truth are faithful for seated and partially occluded bodies; any systematic bias in those reconstructions would directly shift every reported error.","fun_headline_variants_meta":{"raw":{"variants":["ChairPose: full-body pose from pressure alone, unseen chairs","Pressure maps plus chair shape estimate seated pose","Simulation-trained pressure system tracks sitting pose","No wearables: pressure mat reads pose on novel chairs","89 mm pose error from pressure on unseen users and chairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1274,"prompt_tokens":970,"completion_tokens":304,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":586,"tokens_out":304,"duration_ms":4079,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:20:48.392022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a marker-based or multi-view motion-capture system on the same 12 activities across the four chairs, and compare ChairPose's predicted joints to those marker positions for held-out users and chairs; if the mean per-joint error against true marker-based ground truth is substantially above 89.4 mm, especially for arm and head joints, the central claim fails. A second direct test is to place the mattress on a soft, deformable seat and see whether the claimed chair-agnostic accuracy collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the monocular SMPL extraction model used to label the real dataset with pseudo ground truth."},{"cited_title":"Quality and User Experience 8, 1 (2023), 10","cited_arxiv_id":null,"evidence_quote":"provides the physics-based pressure simulation pipeline that ChairPose extends from flat floors to non-planar chair surfaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the ragdoll dynamics used to make synthetic characters interact plausibly with each chair geometry."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the benchmark whose controlled-condition results justify using monocular vision-based pose as ground truth."}],"review_version":1}