{"id":"bf7e4f63-b706-4e3c-b7da-d9ab47ec7141","arxiv_id":"2505.00755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An insole-based system with 35 pressure sensors and an IMU estimates 3D human pose via a Transformer, achieving 49 to 75 mm RMSE in a small five-task experiment.","lead":"This paper builds a low-cost insole with 35 pressure sensors plus an IMU, and uses a Transformer model to estimate a 3D human skeleton from the foot signals. If the results hold, it offers a cheap, privacy-friendly way to monitor posture for rehabilitation and injury prevention.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random frame-level split likely leaks temporal context, so reported RMSE may not reflect generalization to new people or movement sequences.","rationale":"The reader's weakest-assumption analysis correctly identifies the random 8:2 frame-level split as the key threat to the paper's headline numbers. My independent reading confirms that this is the single most load-bearing concern: every other weakness (missing baseline comparison, absent sensor-deployment experiments, omitted dynamic tasks) is secondary to whether the reported RMSE is a truthful estimate of system performance on unseen people. The paper does provide some independent support: the hardware is concretely described, the derivative ablation is a controlled experiment, and the Transformer-vs-LSTM comparison is a reasonable sanity check. But none of that validates generalization across subjects. A leave-one-subject-out or chunk-wise evaluation is a standard, feasible check that would settle whether the model has learned a usable foot-to-skeleton mapping or simply memorized the training distribution. Until that check is run, the appropriate verdict remains conditional, not accept or reject: the method is plausible and not internally contradicted, but the central accuracy evidence is unverified. I therefore see no reason to change the reader's CONDITIONAL verdict.","tokens_in":7920,"tokens_out":5102,"duration_ms":59054,"concrete_test":"Re-run the evaluation with leave-one-subject-out cross-validation (or at minimum split by non-overlapping continuous time windows so that no validation window overlaps any training window by more than 1 second), then recompute the per-task RMSE in Table II. If the average RMSE rises substantially above the reported 48.9–75.2 mm range or exceeds the SolePoser baseline on comparable tasks, the central accuracy claim is not supported by the current evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a Transformer on 35 pressure sensors plus an ankle IMU estimates a 21-joint 3D skeleton with RMSE 48.9–75.2 mm. The supporting evidence is an 8:2 random split over 80 minutes of data from four participants (Section V-D). Because the data are continuous 100 Hz time series, a random frame-level split places validation frames immediately adjacent to training frames from the same movement repetitions. Adjacent frames have nearly identical pressure and IMU inputs and nearly identical skeleton outputs, so the model can memorize the local input–output manifold and achieve low validation error without learning a general foot-to-skeleton mapping. No subject-independent evaluation or split by recording session is reported, and the collected walking, jumping, and hopping tasks are not included in Table II despite being described as part of the dataset. The reported accuracy is therefore not established as representative of new people or of the full daily-activity claim, which is precisely the load-bearing contribution of the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes P2P-Insole, a system that estimates a 21-joint 3D skeleton from foot pressure distribution and motion sensors. The hardware consists of a sub-dollar e-textile insole with 35 pressure sensors and an ankle-mounted IMU; the learning model is an 8-layer Transformer with 8 attention heads trained with MSE loss on synchronized OptiTrack ground truth, pressure, and IMU data. The input stream is augmented with first and second derivatives. Experiments on four participants performing daily movements report RMSE between 48.9 and 75.2 mm across five tasks, plus ablations comparing the Transformer against LSTM and with/without derivative features. The paper claims that the system enables privacy-preserving, camera-free pose estimation for rehabilitation, injury prevention, and health monitoring.","tokens_in":8057,"tokens_out":5518,"duration_ms":56426,"significance":"If the reported accuracy were established, the system would be a practically significant contribution: the sub-dollar hardware cost, the fusion of pressure and IMU data, and the Transformer-with-derivatives design are all reasonable engineering choices, and the ablation against LSTM addresses a relevant design question. The paper also gives a detailed description of the sensor fabrication and data collection pipeline, which is useful for replication. However, the empirical evidence is currently too weak to support the central generalization claim: the evaluation uses only four participants, a random frame-level split, no subject-independent protocol, and no comparison against the directly related SolePoser system that the paper itself lists as achieving sub-70 mm error. No code or data release is mentioned, further limiting reproducibility. The strengths are the low-cost hardware and the clearly stated system architecture; the weakness is that the headline error numbers are not established as representative of new people or of the full daily-activity claim.","major_comments":[{"comment":"The 8:2 random frame-level split described in Section V-D, with training details in Section IV, is not a valid evaluation protocol for 100 Hz time-series data. Because each participant performed continuous movement repetitions and free movement in long sessions, random frame assignment places validation frames temporally adjacent to training frames from the same movement, with nearly identical pressure and IMU inputs and nearly identical skeleton labels. The model can therefore memorize local input-output mappings, and the reported RMSE values in Table II do not estimate performance on new people or new movement sequences. Please report a split by recording session or by participant (leave-one-subject-out), and report per-subject variability.","section":"§V-D and §IV (training protocol)"},{"comment":"The dataset consists of only four participants, and although eight movement classes are collected, Table II reports results only for Stand, Tilt, Bow, Stand and Sit, and Squat; walking, jumping, hopping, and standing on one leg are not evaluated. Without subject-independent evaluation, or at least per-participant results and variance estimates across random seeds, the abstract's claim of robustness and the conclusion's generalization claim are unsupported. The data-collection arithmetic is also inconsistent: eight movements times two minutes plus six minutes of free movement equals 22 minutes, not the stated 20 minutes per participant.","section":"§V-A, §V-B, and Table II"},{"comment":"Table I lists SolePoser as using a similar 32-point insole with an IMU and a two-stream Transformer with sub-70 mm error, and reference [6] is cited, but no quantitative comparison, common benchmark, or discussion is provided. Since the paper's main claim is about pose-estimation accuracy, the lack of comparison against the closest prior art makes it impossible to determine what the new contribution achieves relative to the state of the art. Please add a same-protocol comparison or, if unavailable, clearly state why the numbers are not directly comparable and provide a task-matched baseline.","section":"Table I and reference [6]"},{"comment":"The ablations in Figures 7 and 8 report single mean RMSE values without error bars, per-seed variance, or statistical tests. The statement in Section V-D that the Transformer model significantly improved performance, and the derivative trade-off conclusion, are therefore not supported by the evidence shown. Please report mean plus/minus standard deviation over multiple training runs or use paired tests over subjects.","section":"Figures 7 and 8 and Section V-D"}],"minor_comments":[{"comment":"Equation (1) is described as a voltage divider circuit, but the expression is the standard non-inverting amplifier gain; please reconcile the circuit description with the equation.","section":"Equation (1)"},{"comment":"The notation X in R^{N x d x 2} is unclear: if d is the dimension of the concatenated features, the factor of 2 and the role of N as the number of frames should be defined precisely, especially after the first and second derivatives are added.","section":"Section IV (input notation)"},{"comment":"The preprocessing section states that all data sources were aligned in terms of starting points, ending points, and total number of frames; please state whether data segments were trimmed or padded and how much data was discarded during synchronization.","section":"Section V-C (synchronization)"},{"comment":"The manuscript says the IMU is attached to the subject's ankle but does not specify whether one or both ankles are instrumented; the input description of 82 total dimensions implies two feet, so please clarify the sensor placement.","section":"Section V-A (IMU placement)"},{"comment":"The axis labels and legends in Figures 7 and 8 appear partially machine-generated and are difficult to read; please redraw these figures for publication clarity.","section":"Figures 7 and 8 (readability)"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for publication in its current form because the evaluation protocol undermines the headline accuracy numbers. The hardware and system idea are interesting, and I would be willing to review a revision that adds subject-independent evaluation, variance reporting, and a direct comparison to SolePoser. Please also consider whether four participants is sufficient for the generalization claims; if not, the paper should be reframed as a proof-of-concept with explicit caveats."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible low-cost insole pose-estimation system with a concrete hardware contribution, but the evaluation as written doesn't support the generalization claims. If you work on wearable pose estimation, it's worth a look, but I wouldn't lean on the numbers yet.\n\nWhat's actually new: a sub-dollar e-textile insole with 35 pressure sensors plus an ankle IMU, feeding an 8-layer Transformer that outputs a 21-joint 3D skeleton. The hardware is real and the cost point is interesting. The derivative-feature ablation (first and second derivatives added to the input stream) is a legitimate empirical contribution, with the expected finding that derivatives help for dynamic motions like bowing and squatting but hurt for near-static ones like standing. That trade-off is a useful datapoint.\n\nThe soft spots are mostly about validation. The dataset is four participants, 80 minutes total, and the 8:2 random frame-level split over continuous 100 Hz time series almost certainly leaks temporal context. Adjacent frames from the same movement repetition have nearly identical inputs and nearly identical skeletons, so the model can memorize local structure rather than learn a general foot-to-skeleton mapping. There is no subject-independent evaluation and no session-based split, so the reported 48.9–75.2 mm RMSE is not established for new people.\n\nThe other issue is the missing baseline. The paper cites SolePoser, which already does insole-based full-body pose with under 70 mm RMSE, but never compares against it directly. That's the single most relevant competitor, and the omission makes it hard to know whether this system adds value or just reproduces existing results with cheaper hardware. Also, the paper says it collected walking, jumping, and hopping but Table II only reports standing, tilting, bowing, stand-and-sit, and squatting. The missing tasks matter because those are the more dynamic ones where the claimed robustness should be shown.\n\nNone of this is fatal. The method is not contradicted, and the core idea is sane. But the evidence base is too thin for the accuracy claims to be accepted as they stand. The reader's conditional verdict is right.\n\nWho benefits: researchers working on wearable or privacy-preserving pose estimation, especially those interested in insole-based sensing. It deserves a serious referee, but the referee should require a subject-independent split, a direct SolePoser comparison, and ideally code/data release. I'd send it to peer review with the expectation of major revision, not reject it outright.","headline":"Plausible low-cost insole pose-estimation hardware, but the current evaluation has a temporal leakage problem and no baseline comparison, so the accuracy claims are not yet established.","tokens_in":8605,"tokens_out":1578,"would_cite":false,"duration_ms":18380,"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":"A sub-dollar insole plus ankle motion can reconstruct a 21-joint 3D skeleton.","keywords":["3D human pose estimation","insole pressure sensor","foot pressure distribution","wearable IMU","Transformer","e-textile sensor","privacy-preserving sensing","skeleton prediction"],"falsifier":"Retrain the same architecture on the same data using leave-one-participant-out cross-validation or a time-contiguous split that places whole movement blocks in either training or validation; if RMSE rises well above the reported 75 mm, the central claim of generalizable foot-to-pose estimation fails.","tokens_in":7669,"feed_emoji":"👟","tokens_out":4434,"duration_ms":42643,"temperature":0.7,"pith_summary":"P2P-Insole claims that a pair of embroidered insole sensors, each with 35 pressure points and an ankle-mounted IMU, carries enough information to reconstruct a 21-joint 3D human skeleton across standing, tilting, bowing, sit-to-stand, and squatting. The system reports root-mean-square joint errors between 48.9 mm and 75.2 mm on those tasks, with the largest errors in squatting and the smallest in tilting. If this holds, full-body pose estimation no longer requires cameras, controlled lighting, or expensive motion-capture suits, which matters for rehabilitation, injury prevention, and continuous health monitoring in private spaces. The paper also reports that a Transformer trained on pressure, acceleration, rotation, and their first and second derivatives outperforms an LSTM baseline, with derivatives helping most on large-movement tasks.","feed_headline":"A $1 insole reconstructs your 3D skeleton","feed_subtitle":"Foot pressure plus ankle motion locates 21 joints with 49–75 mm error, no camera required.","key_machinery":"The load-bearing machinery is an 8-layer Transformer encoder over time-series frames, where each input frame concatenates 35 pressure values, 3 acceleration axes, and 3 rotation axes from each foot, together with the first and second derivatives of those signals. The derivatives turn static pressure images into motion-sensitive temporal features, which is what lets a foot-only sensor infer upper-body joints such as the head and arms. The 35-sensor insole geometry, the 100 Hz sampling, and the ankle IMU placement together define the input stream; the paper argues that this stream, not the model alone, is what makes sub-8 cm full-body estimates possible.","core_discovery":"The paper claims that plantar pressure distribution plus foot acceleration and rotation is a sufficient signal for estimating full-body 3D pose in a set of daily movements. A custom e-textile insole (under USD 1 per unit, with 35 piezoresistive Velostat-based sensors) captures pressure at 100 Hz, an integrated IMU adds 3-axis acceleration and rotation, and an 8-layer Transformer with 8 attention heads maps an 82-dimensional per-frame input to a 63-dimensional skeleton output. The model is trained on eighty minutes of synchronized data from four participants wearing OptiTrack suits, with first and second time-derivatives of each input channel appended. Reported RMSE is 48.9 mm for tilting, 52.6 mm for standing, 63.7 mm for bowing, 70.1 mm for standing and sitting, and 75.2 mm for squatting, and the Transformer outperforms an LSTM baseline across all tasks.","pith_inferences":["The reported accuracy is measured on a random 8:2 frame split from four participants; a stricter leave-one-participant-out evaluation would reveal how much of the learned mapping is person-specific and how much generalizes to new users.","The derivative trade-off suggests a task-adaptive input selector: include derivatives for dynamic movements and exclude them for quasi-static poses, a design the paper does not test.","If the foot-signal mapping generalizes, the same input stream could support contact-based fall prediction and gait analysis without a skeleton label, since the Transformer's temporal features already encode balance and weight shift.","Sensor-count reduction based on the reported sensor-position/error relationship could identify a minimal subset of pressure points that preserves accuracy, which the conclusions flag as future work."],"forward_implications":["A camera-free pose-estimation pipeline becomes feasible in bedrooms, bathrooms, and workplaces where cameras are unacceptable.","The reported error distribution implies that lower-body joints are estimated much more tightly than the head and arms; spine median error is about 6 mm while head median error is about 56 mm.","Adding time derivatives to the input is beneficial for large-amplitude motions such as bowing and squatting but can hurt small-movement accuracy such as standing and tilting, so input augmentation should be task-dependent.","Because the insole costs under USD 1 and uses an off-the-shelf microcontroller, the hardware could be mass-produced for continuous daily monitoring rather than lab sessions."],"supporting_citations":[{"why":"The closest prior work estimating full-body pose from a pair of insoles, used as the accuracy comparison point.","marker":"[6]"},{"why":"Extends insole pressure to dual-modal pose estimation and is the direct methodological predecessor.","marker":"[22]"},{"why":"Shows IMU-only pose estimation, whose drift and discomfort problems motivate adding the pressure modality.","marker":"[7]"},{"why":"Supplies the argument that contact and pressure information improves pose estimation in human-scene interaction.","marker":"[10]"},{"why":"Demonstrates 3D pose inference from tactile signals, establishing the pressure-to-pose premise for other body locations.","marker":"[20]"},{"why":"Introduces spatial-aware ground tactility for monocular pose estimation, supporting the use of foot pressure as an auxiliary signal.","marker":"[19]"},{"why":"Shows that full-body pressure garments can infer 3D poses, serving as a reference point for wearable pressure-based pose learning.","marker":"[24]"},{"why":"The authors' prior insole-based fall-detection work, providing the sensor platform lineage for the current system.","marker":"[23]"}],"fun_headline_variants":["A $1 insole turns foot pressure into full 3D body pose","Foot pressure and IMU on a $1 insole estimate your 3D skeleton","Cheap insole reads pressure to map 3D pose without cameras","Under-$1 insole uses foot motion to reconstruct whole-body skeleton","Sub-$1 insole with pressure sensors and IMU predicts 3D pose"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a random 8:2 split of consecutive frames from four people's movement sessions measures true accuracy; if neighboring frames from the same movement appear in both training and validation, the model may memorize pressure patterns instead of learning a general foot-to-skeleton mapping.","fun_headline_variants_meta":{"raw":{"variants":["A $1 insole turns foot pressure into full 3D body pose","Foot pressure and IMU on a $1 insole estimate your 3D skeleton","Cheap insole reads pressure to map 3D pose without cameras","Under-$1 insole uses foot motion to reconstruct whole-body skeleton","Sub-$1 insole with pressure sensors and IMU predicts 3D pose"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1611,"prompt_tokens":905,"completion_tokens":706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":605}},"tokens_in":521,"tokens_out":706,"duration_ms":7048,"temperature":1.0,"reasoning_tokens":605,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:42:16.355298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same architecture on the same data using leave-one-participant-out cross-validation or a time-contiguous split that places whole movement blocks in either training or validation; if RMSE rises well above the reported 75 mm, the central claim of generalizable foot-to-pose estimation fails.","supporting_citations":[{"cited_title":"Soleposer: Full body pose estimation using a single pair of insole sensor,","cited_arxiv_id":null,"evidence_quote":"The closest prior work estimating full-body pose from a pair of insoles, used as the accuracy comparison point."},{"cited_title":"Dual-modal 3d human pose estimation using insole foot pressure sensors,","cited_arxiv_id":null,"evidence_quote":"Extends insole pressure to dual-modal pose estimation and is the direct methodological predecessor."},{"cited_title":"Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time,","cited_arxiv_id":null,"evidence_quote":"Shows IMU-only pose estimation, whose drift and discomfort problems motivate adding the pressure modality."},{"cited_title":"Populating 3d scenes by learning human-scene interaction,","cited_arxiv_id":null,"evidence_quote":"Supplies the argument that contact and pressure information improves pose estimation in human-scene interaction."},{"cited_title":"Intelligent carpet: Inferring 3d human pose from tactile signals,","cited_arxiv_id":null,"evidence_quote":"Demonstrates 3D pose inference from tactile signals, establishing the pressure-to-pose premise for other body locations."},{"cited_title":"Satpose: Improving monocular 3d pose estimation with spatial-aware ground tactility,","cited_arxiv_id":null,"evidence_quote":"Introduces spatial-aware ground tactility for monocular pose estimation, supporting the use of foot pressure as an auxiliary signal."},{"cited_title":"Learn to infer human poses using a full-body pressure sensing garment,","cited_arxiv_id":null,"evidence_quote":"Shows that full-body pressure garments can infer 3D poses, serving as a reference point for wearable pressure-based pose learning."},{"cited_title":"Pifall: A pressure insole-based fall detection system for the elderly using resnet3d,","cited_arxiv_id":null,"evidence_quote":"The authors' prior insole-based fall-detection work, providing the sensor platform lineage for the current system."}],"review_version":1}