Source-linked AI summary
FISICA: A Deployed Service for Plantar-Pressure and Posture Assessment with Ontology-Grounded Recommendation
Juhwan Song, Heejung Kim, Juntae Noh, Jonghak Ryu, Huiju Park, Junseong Lee, Dohyeon Ahn, Byungwoo Jo
TL;DR
FISICA tackles the accessibility and cost of conventional standing-posture and foot-loading assessment by deploying a production service that combines a pressure scale with two RGB photographs. It introduces closed-loop avatar measurement using a sampling-invariant spinal metric, while rules control recommendations and the language model only explains stored decisions. The service reports 7.2° separation between normal and kyphotic records versus 0.9° for a single-joint formulation, with sub-second rule evaluation and measured production latency across its features.
Problem
Standing-posture and foot-loading tests usually require costly, hard-to-access radiography or laboratory assessment.
Method
FISICA combines a pressure-sensing scale, RGB capture, analysis, versioned storage, rule-based recommendation, and an avatar measured with the same function as the subject.
Results
7.2° separates a normal from a kyphotic record versus 0.9° for the replaced single-joint formulation, while production reports 0.023 s general APIs, 0.45 s plantar-pressure analysis, and 2.16 s to 2.26 s recommendations.
Takeaways & Limitations
The deployed service provides traceable ranked shoe and exercise candidates from stored measurements, with every recommendation controlled by rules and explained by a language model.
Takeaways & Limitations
The posture comparison uses two records rather than a validation cohort and has no radiographic reference; published service-latency comparisons are not head-to-head tests.
Abstract
from arXiv · showhide
FISICA is a body-assessment and recommendation service running in production. One standing session with two photographs returns foot-loading measures, posture coordinates, a driven 3D avatar, a visual report, and ranked shoe and exercise candidates. Measurement comes from a purpose-built scale carrying 634 force-sensitive elements on a 1 cm grid and four load cells, and a rule-based evaluator controls every recommendation while a language model only explains the stored result. The method contribution is the avatar. Instead of mapping a measured angle onto a rig through a tuned gain, we measure the avatar with the same function used on the subject and solve until the two agree, on a sampling-invariant spinal metric that separated a normal from a kyphotic record by 7.2 degrees against 0.9 degrees for a single-joint formulation. In production, general APIs respond at a 0.023 s median, plantar-pressure analysis at 0.45 s, and recommendation at 2.16 s to 2.26 s with the rule-based portion under one second in every trial. The served keypoint graph reaches 0.960 PCK@0.2 on public data, and the catalog holds 699 shoes with 10,500 typed facts. An approved study supplies the radiographic reference for the validation still ahead.
1 Introduction
FISICA addresses the cost and access barriers of conventional standing-posture and foot-loading assessment with a production service built around pressure sensing and RGB photography. Its main methodological contribution is a self-measured, sampling-invariant avatar that supports traceable recommendations within a wellness scope.
- Radiography or laboratory testing is usually required for standing-posture and foot-loading assessment, whereas FISICA uses a pressure-sensing scale and two RGB photographs.
- One standing session produces plantar-pressure measures, foot classification, posture indicators, a driven avatar, ranked shoes, and safety-filtered exercises.Every output is stored with the measurement record and the model, rule, and catalog versions that produced it.
- FISICA measures its avatar with the same function applied to the subject and solves until the two measurements agree, rather than using a hand-tuned mapping.The spinal metric is designed to remain invariant when the spine is sampled at different densities.
- 7.2° separates a normal from a kyphotic record, versus 0.9° for the replaced single-joint formulation.
- Recommendation scores are catalog ranks rather than probabilities of benefit, and posture indicators and class labels are analysis outputs rather than diagnoses.FISICA is explicitly framed as a fitness and wellness service.
2 Related work
Prior work establishes photographic posture measures, pressure-platform reductions, and pose-estimation pipelines, but FISICA focuses on validating the connection between measured findings and a running service. Its avatar construction closes that connection through same-function measurement and closed-loop agreement.
- Human pose estimation has progressed from part-based models to direct regression and top-down detectors, with convolutional and transformer backbones supporting different accuracy-speed trade-offs.FISICA uses such a model as a component and evaluates its exported graph under the COCO keypoint protocol.
- Photographic posture measures and the craniovertebral angle are established, but their reliability depends on camera placement, landmark definition, and rater technique.
- FISICA measures its avatar with the same function used on the subject and solves until the measured values agree, rather than relying on a hand-tuned mapping.The construction requires a metric whose comparison is well posed across different spine sampling densities.
- Pressure platforms commonly reduce load distributions and footprint geometry through the arch index and its anatomical thirds; FISICA adopts that reduction directly.
- A service boundary in which a language model only explains a rule-evaluator decision addresses versioning, provenance, and technical-debt concerns.
3 System architecture
FISICA integrates sensing, capture, analysis, storage, recommendation, visualization, and reporting as one versioned system. Its scale separates spatial loading from total mass and derives physical foot dimensions without camera calibration or external reference objects.
- The system combines a pressure-sensing scale, frontal and lateral RGB capture, measurement and analysis services, a versioned domain graph, evaluators, and an avatar client.The components are operated as one system under one team, preserving provenance from sensor frame to ranked product.
- The exploded scale integrates pressure-sensing layers and PCB hardware for spatial loading, alongside four load cells that provide total body mass.
- Measurement records retain capture, application, firmware, and payload provenance, while analysis adds derived observations without overwriting source data.Recommendations receive compact observation projections rather than raw images or the full pressure matrix.
- A fixed physical sensing pitch lets FISICA report foot dimensions in millimetres without a camera reference or calibration target.
- The device uses a layered glass, silicone, conductive, PCB, and load-cell construction designed to spread load and keep array response repeatable.
- A separate pressure path and mass path do not share a converter, allowing spatial distribution and absolute mass to be checked independently.The per-cell sum can be checked against the array sum as a validity test.
- Certification and intellectual-property provisions frame the scale as a shipping product rather than a laboratory rig.Service-level measurements were taken under that product condition.
- The architecture includes thirteen microservices spanning gateway, identity, measurement, inference, storage, audit, recommendation, and conversational-agent functions.
4 Plantar-pressure analysis
FISICA converts a fixed-grid pressure frame into foot dimensions, regional loading measures, and a nine-class label. Its geometry pipeline compensates for coarse sampling and missing toe activations without camera scaling or user input.
- Frame format: 645-byte frames encode a 22 × 29 grid at 1 cm pitch, with 634 sensing points divided into two 11 × 29 foot regions.The frame also carries body mass, battery, and terminator fields.
- Foot geometry: Foot length combines principal-axis extent, cell-diagonal dilation, metatarsal-ball localization, and a capped template estimate.The system reports the observed extent unless the template-completed estimate is larger within the fixed cap.
- Foot geometry: A row-profile test detects whether toes registered, enabling millimetre foot-length recovery from the physical grid without camera reference or user input.The test uses forward gap, toe-to-ball width ratio, and narrowing between them.
- Regional loading: Forefoot, midfoot, and rearfoot comprise the leading 28 %, middle 32 %, and trailing 40 % of each foot’s bounding rows.Band loads are percentages of each foot’s total; symmetry is the ratio of half-mat sums, and the centroid is pressure-weighted and normalized to the unit square.
- Regional loading: The centroid is a single static standing-frame measurement, not a time-varying centre-of-pressure trajectory.This defines the temporal scope of the reported balance-related measure.
- Classification: A nine-way convolutional classifier reads the pressure grid and reports its top three classes from a softmax distribution.The labels include foot types and posture-related categories such as lordosis, kyphosis, scoliosis, and pelvic torsion.
5 Posture estimation
FISICA estimates posture from frontal and lateral RGB captures using a deployed 2D keypoint graph, with an optional high-compute 3D mode. The evaluated 2D path supplies coordinates for geometric extraction and avatar driving, while its public-data assessment uses scale-aware reporting.
- Two modes: The 2D ONNX graph returns joint coordinates, depth, and a service-specific C7 landmark, with face regions blurred in returned imagery.These outputs feed geometric extraction and the avatar.
- Evaluation: The held-out evaluation contains 120 COCO validation images and 168 annotated person instances, including 83 single-person images.The split is independent of the approved collection described for later validation.
- Evaluation: The COCO evaluation does not match FISICA’s operating condition, so error is reported against subject scale rather than averaged over the scale distribution.The product condition is one standing subject filling the frame at a fixed distance.
- Metrics: PCK uses the ground-truth person bounding-box diagonal as its normalizer and counts every visible annotated keypoint.Unmatched instances count as failures, preventing detection misses from inflating the value.
- Metrics: Classification metrics are undefined because the model emits a fixed 17-keypoint vector without a true-negative or false-negative keypoint population.The paper therefore does not report accuracy, precision, or recall for this task.
6 Measured accuracy and cost of the vision graph
The deployed vision graph is evaluated on public COCO data for accuracy, latency, dispersion, and memory. Accuracy is strong on the held-out split, while derived-angle reliability depends on landmark scale and inference cost depends on candidate-instance count.
- Keypoint accuracy: 0.960 PCK@0.2 was achieved by the deployed configuration, which matched every ground-truth instance on the held-out split.The split contains 120 public COCO images and 168 person instances.
- Keypoint accuracy: 0.847 PCK@0.2 and 0.499 AP were achieved by BlazePose on the 83 single-person images, versus 0.953 and 0.616 for the deployed graph.The authors include BlazePose as a practical alternative rather than a head-to-head claim.
- Keypoint accuracy: Head landmarks have normalized errors of 0.016–0.033, shoulders 0.035–0.046, and hips 0.056–0.065.The larger hip error reflects inferred joint centers; pelvic indices are therefore treated more conservatively than shoulder indices.
- Latency and memory: 62.8 ms, 49.4 ms, and 54.2 ms are the three configuration medians, differing by at most 14 ms under the interleaved protocol.Input resolution and numeric precision changed accuracy without materially changing served latency.
- Latency and memory: 66.3 ms for one instance rose to 119.2 ms above twenty instances, while the p99 was 2.7 times the p50.The deployed service keeps only the first detection, so the multi-person tail belongs to benchmark images rather than deployment.
- Latency and memory: 1659 MiB of host resident memory and 1096 MiB of device memory above idle were measured separately for capacity planning.The host peak largely reflects the decoded benchmark image set; half-precision weights account for 154 MiB at load time.
- Derived angles and subject scale: 2.92° and 4.03° are the median shoulder- and pelvic-tilt errors after excluding sign reversals and 20–90° errors.Sign reversals affect 3.8% of shoulder-tilt and 6.9% of pelvic-tilt instances; the authors recommend detecting reversals and abstaining.
- Derived angles and subject scale: 5.21° below 40 px of shoulder separation fell to 1.91° above 80 px, where reversal rates changed from 9.4% to zero.The fixed-distance, single-subject capture condition falls in the largest scale bin, so trend-by-scale is more informative than marginal averages.
7 Embodied posture visualization
FISICA drives a rigged avatar by measuring it with the same posture functions applied to the subject and solving until the measurements agree. A sampling-invariant two-line spinal metric enables this closed loop and preserves the observed thoracic difference between records.
- 7 Embodied posture visualization: The pipeline estimates keypoints, extracts posture parameters in an anatomical frame, and drives the avatar until its measurements match the subject’s.Keypoints are lifted into a body-fixed frame, normalized by torso length, and processed through estimation, extraction, and embodiment stages.
- 7.2 The sagittal bend metric: Sampling dependence makes the convexity index unsuitable for comparing five-point human spines with densely sampled avatar spines.Its maximum-deviation formulation reports larger values for densely sampled curves of the same shape.
- 7.2 The sagittal bend metric: 0.9◦ separates the normal and kyphotic records with the single-joint bend, which captures curvature at only one vertebral node.The records measure 27.7◦ and 28.6◦, and the thoracic apex is not fixed across subjects.
- 7.2 The sagittal bend metric: The adopted metric traverses the spine by arc-length fraction and compares fitted lower and upper lines in sagittal or coronal planes.Thoracic and lumbar channels use specified fractional intervals, separating posture shape from lateral deviation and trunk lean.
- 7.2 The sagittal bend metric: 7.2◦ separates the two records with the two-line thoracic bend, compared with 0.9◦ for the single-joint formulation.The metric is sampling-invariant, apex-agnostic, and lean-invariant, making closed-loop matching possible.
- 7.5 Two-record comparison: The avatar transfers the thoracic difference at about 7◦ versus 7.2◦ on the subject, while its absolute value remains about 4◦ higher in both conditions.The two records were used to build and check the metric, not as a validation cohort or radiographic comparison.
8 Ontology-grounded recommendation
FISICA uses a typed ontology, deterministic rules, and catalog facts to produce traceable shoe and exercise recommendations. The language model explains a stored decision but does not determine the ranking.
- 8.1 Knowledge representation: The ontology stores typed concepts, relations, mapping rules, item facts, derived features, evidence, claim policy, and separate user assertions.Typed need identifiers restrict which concepts can enter evaluation.
- 8 Ontology-grounded recommendation: The development recommendation layer maps structured observations to footwear and exercises through a PostgreSQL graph query followed by Java evaluation.The evaluator checks thresholds and composite conditions, applies hard exclusions before scoring, and ranks candidates.
- 8.1 Knowledge representation: MAY_INDICATE and SHADOW_ONLY relations may support review but cannot contribute to active scores, so posture-to-footwear hypotheses are excluded from shoe ranking.This boundary reflects the posture channel’s scale-dependent accuracy and possible sign reversals.
- 8.2 Footwear catalog and scoring: The catalog contains 699 running-shoe models, 10,500 typed product facts, 7,413 feature assertions, 19 mapping rules, and 8 evidence records.Missing laboratory values remain UNKNOWN rather than being replaced by zero or a catalog mean.
- 8.2 Footwear catalog and scoring: Coverage enters the footwear score through qf, while thresholds in the development scorer are engineering conditions rather than validated clinical cut-points.The evaluator also applies requested-brand filters, avoidances, product issues, and bounded personalization terms.
- 8.4 Explanation and safety: The language model converts a compact, evidence-bounded brief into user language, while a response checker rejects changed rankings, unsupported entities, and diagnostic or treatment claims.The stored graph-and-rule decision remains the recommendation authority.
9 Service performance
FISICA reports service-level performance across production usage, scenario tests, and controlled experiments. Recommendation latency is dominated by language-model generation, while several measurement paths remain substantially faster.
- 9 Service performance: 0.023 s is the production median for general APIs across 25,492 requests, while agent conversations have a 2.51 s median.The general-API p95 is 0.249 s; agent conversations have a 6.93 s p95.
- 9.3 Recommendation flow: Recommendation processing consists of admission, language-model planning, evaluator matching, language-model generation, and answer streaming.Figure 13 reports the median cost of each stage; stages 1 to 3 are rule-based and stages 4 to 5 generate language.
- 9.3 Recommendation flow: 0.78 s for shoes and 0.84 s for exercises are the production medians of the rule-based recommendation portion, which stayed under one second in every trial.End-to-end exercise recommendation has a 2.26 s median, 3.55 s standard deviation, and 13.26 s p95.
- 9.4 ML inference in real usage: 37.4 s to 46.4 s reflects cold model loading for 3D mode, whereas warm calls take 3.0 s to 9.5 s.The first call of an infrequent burst includes lazy model loading, so the cold figure is not GPU capability.
- 9.1 Measurement services: 0.45 s is the production median for plantar-pressure analysis, with a 1.52 s p95.This path is reported as substantially faster than recommendation flow.
10 Approved reference-data collection
The approved protocol establishes a prospective reference resource pairing posture photographs with spinal radiographs for evaluating image-derived posture measures. Its analysis labels remain proposed and require definitions, qualified readers, and disagreement procedures before validation.
- Approval and data scope: Protocol P01-202512-01-047 was approved under institutional review-board oversight for collecting posture photographs, spinal radiographs, and basic participant information.The study was classified as minimal risk and requires written consent.
- Approval and data scope: The analysis dataset excludes direct identifiers, and no participant-level result from this protocol appears in the manuscript.
- Pairing and annotation: Frontal and lateral photographs are paired with corresponding radiographs at the record level, without treating RGB and X-ray pixels as a shared coordinate system.Annotators use one surface-landmark definition across RGB views and a separate manual for radiographic reference points.
- Validation prerequisites: The proposed posture and foot labels require defined references, reader qualifications, and disagreement procedures before model validation.The listed labels include normal posture, foot deformity, flat feet, lordosis, kyphosis, scoliosis, and pelvic tilt; they are not current-service diagnoses.
- Validation prerequisites: Prospective validation will quantify agreement between image-derived posture coordinates and angles and radiographic landmarks using a prespecified protocol.The plan includes cohort counts, exclusions, capture geometry, de-identification, blinded readers, inter-reader agreement, and statistical testing.
11 Conclusion
FISICA operates as an integrated, versioned service whose rules control recommendation filtering and ranking while the language model explains stored results. Its deployed measurements show low service latency, public-data pose accuracy, scalable GPU execution, and traceable catalog coverage.
- Conclusion: A pressure frame and two photographs become a versioned measurement record containing foot and posture findings, a driven avatar, a report, and shoe and exercise candidates.The service links these outputs across one pipeline rather than presenting separate model results.
- Conclusion: Rules control recommendation filtering and ranking, while the language model selects approved tools and explains stored results.
- Conclusion: 0.45 s plantar-pressure analysis and 0.023 s general APIs accompany 2.16 s shoe and 2.26 s exercise recommendations, with rule-based processing under one second in every trial.The reported recommendation comparison is against published tool-calling agent figures of 2.8 s to 5.9 s after optimization.
- Conclusion: 0.960 PCK@0.2 and 0.633 AP are reported for the deployed keypoint graph, while the catalog contains 699 shoes and 10 500 typed facts.Every ranking carries a trace back to the measurement that produced it.
12 Future work
Future work targets deployment measurement, safer derived-angle handling, more complete avatar channels, deeper catalog data, and reference-based validation. The proposed changes address accuracy reporting, abstention, instrumentation, item metadata, and radiographic comparison.
- 12 Future work: Native-resolution full-precision vision export would provide 0.097 AP and 0.081 mean OKS while keeping three configurations within 14 ms of one another.The stated comparison places the configurations within the current latency budget.
- 12 Future work: A scale gate and reversal check would withhold tilt indices when left-right reversals occur, converting that failure mode into abstention.The reversals disappear above 80 px of landmark separation.
- 12 Future work: Avatar work would unify craniovertebral angle, round-shoulder protraction, and shoulder level definitions across channels.Per-channel solver residuals would help resolve the thoracic offset and confirm lumbar range and saturation.
- 12 Future work: Instrumentation should log pose request modes and keep the high-compute model resident or warmed so first-call loading does not affect served latency.
- 12 Future work: Catalog expansion would add inventory state, size-specific internal dimensions, and the relation for the remaining exercise item.
- 12 Future work: Protocol P01-202512-01-047 will quantify agreement between image-derived posture angles and radiographic landmarks.
A Supplementary figures
The supplementary figures expose the deployed topology and a ranked-candidate page, connecting infrastructure, recommendation evidence, and released catalog facts. Together they show how a measurement becomes an inspectable recommendation.
- Figure 16: Deployed topology: Figure 16 depicts an AWS edge tier connected by WireGuard to production and development k3s clusters running the service stack.The production cluster has two GPU worker nodes and thirteen microservices, while development uses one node.
- Figure 17: Candidate page: Figure 17 shows the top-ranked candidate’s fit, stock, need-against-item radar, match score, dimension breakdown, and grouped catalog facts.The page groups facts into dynamic stability, static stability, mobility, and fit.
- Figure 17: Candidate page: Every displayed recommendation number traces to a catalog fact identifier and derivation rule, linking the item profile to released laboratory attributes.