Source-linked AI summary
More Motion Is Not Always Better Motion: Corpus Composition Governs Whether Augmentation Helps SMPL-Based Parkinsonian Gait Severity Estimation
Michael Caiola, Andrew C. Weitz
TL;DR
The paper asks whether motion augmentation improves SMPL-based Parkinsonian gait-severity estimation and tests frozen MotionAGFormer encoders across lifting corpora and representation variants. The final ensemble reaches macro-F1 0.58 on a hidden multi-site test set, while gains depend on walking-speed contrast rather than volume, and several external or representation modifications fail to help.
Problem
The study addresses gait-severity grading from SMPL motion when labeled CARE-PD data are small and imbalanced and hidden evaluation sites differ from training sites.
Method
The paper evaluates frozen MotionAGFormer encoders and ensembles while varying inertial lifting-corpus task composition, external sources, and learned-representation modifications.
Results
Macro-F1 0.58 is achieved on the hidden multi-site test set, versus 0.51 without outside motion and 0.53 for the strongest single member.
Takeaways & Limitations
Augmentation helps when added motion carries a walking-speed contrast, whereas pools lacking that contrast, exact synthetic motion, web-video reconstructions, and representation modifications do not help.
Takeaways & Limitations
Offline leave-one-cohort-out comparisons did not reliably predict hidden-test movement, and the volume comparisons cover only a roughly sixfold range.
Abstract
from arXiv · showhide
We grade MDS-UPDRS gait severity from SMPL motion using three frozen MotionAGFormer encoders as featurizers, reaching macro-F1 0.58 on a hidden, multi-site test set. Because the system's members differ only in their lifting corpus, evaluating encoders singly on that test set isolates what that corpus contributes. Six pools drawn from one inertial dataset, varying only in which walking tasks they include, score between 0.32 and 0.53, and just one of them beats the 0.51 of an encoder given no outside motion at all. What separates them is not how much data they hold but whether they carry a contrast in walking speed, the variation this representation appears to depend : a further pool adding a third collection site at fixed task composition does worse still. The same rule explains why exact synthetic motion and monocularly reconstructed web video both fail to help. Modifying the learned representation itself, rather than the corpus behind it, cost every variant that attempted it.
1 Introduction
The paper studies gait-severity grading from SMPL motion under limited, imbalanced labels and hidden multi-site evaluation. It finds that augmentation helps only when added motion supplies the representation’s relevant walking-speed contrast, while corpus source and evaluation design also matter.
- Task and motivation: The task grades MDS-UPDRS 3.10 gait severity from a single canonicalized SMPL walk using only body motion.The ordinal labels are 0, 1, 2, and 3; video, appearance, and clinical metadata are excluded.
- Task and motivation: Only 2950 of CARE-PD’s 8477 walks carry gait grades, leaving a small and imbalanced labeled pool.CARE-PD spans 362 subjects across nine cohorts, but only four cohorts are labeled for this task.
- Task and motivation: Outside motion is attractive for augmentation, but optical capture, inertial sensing, synthesis, and monocular recovery are not interchangeable after conversion to SMPL.Each source introduces its own reconstruction process, so identical output parameters do not guarantee equivalent training value.
- Central finding: An inertial corpus helped when its pool included hurried and self-paced walking, but hurt when that speed contrast was absent.The paper identifies walking-speed variation as the contrast this representation appears to depend on.
- Study scope: The study reports augmentation experiments, an ensemble comparison, six representation-modification variants that all lost, and three external sources that failed to transfer.Offline and hidden-test results also diverged because unseen cohorts differed in capture and reconstruction from training data.
2 Materials and Methods
The methods construct lifting corpora and frozen MotionAGFormer ensembles from multi-cohort motion data, then evaluate augmentation, feature alternatives, alignment, and generalization under leave-one-cohort-out and hidden-test protocols.
- Data sources: CARE-PD contributes four labeled cohorts covering 2950 walks and five unlabeled cohorts, while outside lifting sources include synthetic BEDLAM and inertial WearGait-PD motion.WearGait-PD uses body-worn inertial sensors across three sites, with the main results generally using the two Johns Hopkins sites.
- Lifting corpora: WearGait sequences are converted to pose-only SMPL motion by inverse kinematics, then expanded through two temporal downsampling rates and two camera azimuths.Translation is set to zero because inertial data alone cannot recover it, and the two-site pool yields 698 training sequences.
- Encoder and features: Four lifting encoders isolate corpus effects by varying only the added corpus or schedule, while MotionAGFormer produces 512-dimensional walk descriptors from averaged 81-frame, two-view embeddings.The encoder is frozen and receives no severity labels; supervision is applied only to the MLP head.
- Lifting corpora: The WearGait pools differ in walking-task composition, subject coverage, and segmentation; a seventh pool adds a third site while preserving task proportions and the processing protocol.The six primary pools include whole-trial and short-clip variants, including self-paced-only and mixed-task compositions.
- Encoder and features: Each member standardizes features using its own training statistics and predicts four classes with a class-weighted MLP; ensemble members differ by backbone size and seed.At inference, member softmax outputs are averaged before taking the arg-max.
- Feature baselines: The study compares learned embeddings with 434 joint-rotation statistics and 26 clinical spatiotemporal gait parameters over the same SMPL motion.The hand-crafted pipelines include rotational variability and spectral content, plus cadence, step length, stride-time variability, and speed.
- Ensemble design: Backbone compositions span four MotionAGFormer sizes and ensembles of two, three, or four members; the released system uses one small and two large encoders weighted 0.4/0.3/0.3.The small member is upweighted because it uses the WearGait-augmented corpus.
- Test-time alignment: Transductive alignment recenters test features globally or within site-correlated fingerprint groups, with fallback to training statistics for uncertain assignments.The method addresses large site-driven feature variance in a test set whose sites are absent from training.
3 Results
WearGait augmentation helped only when its task composition included hurried-paced walking, while larger or differently sourced motion did not reliably transfer. The final frozen-encoder ensemble reached macro-F1 0.58, but modifying the representation or replacing its learned features reduced hidden-test performance.
- Lifting-corpus composition: WearGait augmentation improved the unaugmented encoder on the hidden test, whereas BEDLAM remained worst and the longer WearGait schedule ranked second worst.The hidden test differs partly from offline validation: the longer schedule was best offline but performed poorly on the held-out multi-site test.
- Lifting-corpus composition: Only one of six WearGait pools exceeded the 0.51 unaugmented baseline; among whole-trial pools, higher hurried-pace share matched higher hidden-test macro-F1, unlike corpus size.The largest pool contained 2.5 times the frames and 35 more subjects than the selected pool yet scored 0.06 lower.
- Lifting-corpus composition: 0.50 versus 0.58: adding a third collection site at fixed task proportions lowered the three-member ensemble’s hidden-test macro-F1.The corresponding LOCO macro-F1 fell from 0.370 to 0.356, while class-0 recall fell from 0.625 to 0.474.
- Lifting-corpus composition: Speed matching reduced 0-versus-1 discrimination from 0.63 to 0.47 AUC, suggesting that this representation depends on walking-speed contrast.The three other representations tested under the same speed-matched condition also reached chance.
- Ensemble composition: The small 26×64 encoder was load-bearing: pairing it with larger members and a third seed helped, while a fourth member cost roughly 0.02.The reported composition comparisons indicate that capacity alone did not explain the ranking.
- Ensemble composition: The final three-member ensemble reached macro-F1 0.58, with precision 0.65, recall 0.55 and accuracy 0.53.Changing ensemble weights produced identical values on these metrics; quadratic weighted kappa differed only slightly, 0.41 versus 0.40.
- External sources that did not transfer: Feature-space modifications scored 0.49, 0.51 or 0.52 against the 0.58 baseline, while full unfreezing scored 0.40 and 0.35.Every system that left the features untouched returned 0.58, and larger representation changes were associated with lower scores.
- External sources that did not transfer: Hand-crafted replacements scored 0.41 for 434 joint-rotation statistics and 0.34 for 26 spatiotemporal gait parameters, versus 0.58 for the learned-feature ensemble.The hand-crafted pipelines therefore underperformed the frozen learned representation on the hidden test.
4 Discussion
Augmentation helps when added motion contains the speed contrast this representation appears to depend on, whereas changing reconstruction sources, sites, or the learned representation can undermine transfer. Offline validation is limited because it does not reliably predict hidden-test movement and cannot meaningfully evaluate class 3.
- What governs whether augmentation helps: 0.32–0.53: the six pools span this hidden-test range against 0.51 for adding nothing, while the three-site pool performs worse than composition variants.The pools share one corpus and conversion pipeline; the three-site comparison adds 58 subjects at fixed task composition.
- What governs whether augmentation helps: BEDLAM hurts despite exact SMPL parameters and greater size, behaving like pools without a speed contrast.This supports composition and variation, rather than volume or parameter exactness alone, as the relevant distinction.
- What governs whether augmentation helps: Monocular web-video reconstruction cannot provide clean speed contrast because uncalibrated-camera scale errors propagate to length- and speed-based quantities.Scale normalization or scale-invariant features are needed before web video can serve as training data.
- Limitations: CARE-PD requires generalization across reconstruction methods as well as sites, and splits over four cohorts predict the hidden test poorly.Its cohorts use different capture modalities and fitting pipelines.
- Why the representation resists improvement: 0.35–0.52: all six variants that altered the frozen representation scored within this range, with larger alterations producing larger losses.The modifications recentered features, changed camera geometry, or trained the encoder.
- Why the representation resists improvement: Class 3 is the likeliest casualty of representation perturbation because walking-aid information is absent from SMPL and the class appears at a feature-space extreme.This remains a likelihood rather than a directly measurable attribution from aggregate scores.
- Limitations: Leave-one-cohort-out cannot measure class 3, so offline macro-F1 covers classes 0–2 while hidden-test macro-F1 covers all four.Class 3 has only 44 walks from seven subjects in two labeled cohorts.
- Limitations: The 1.77-grade reconstruction-variance result uses only eight paired walks, establishing possible movement beyond one grade without precisely estimating effect size.Hidden-test comparisons are also single measurements, so small adjacent-row differences are unresolved.
5 Conclusion
The system estimates MDS-UPDRS gait severity from SMPL motion using frozen MotionAGFormer encoders and reaches 0.58 macro-F1 on a hidden multi-site test. The paper finds that augmentation quality depends on motion content and pipeline placement, while modifying the learned representation consistently harms performance.
- What to add: Speed contrast, not volume, predicts whether added motion helps across the tested corpora.Pools without speed contrast score below adding nothing, including larger or exact synthetic motion.
- What not to touch: Every variant that modified the frozen representation lost, with greater modification associated with greater loss.Offline evaluation did not reveal the damage in advance.
- Availability: The final system, checkpoints, and corpus-building scripts are available at the project repository, while source datasets remain governed by their original authors’ terms.The paper identifies CARE-PD, WearGait-PD, BEDLAM, and GAVD as externally provided datasets.