More Motion Is Not Always Better Motion: Corpus Composition Governs Whether Augmentation Helps SMPL-Based Parkinsonian Gait Severity Estimation

Michael Caiola; Andrew C. Weitz

cs.CVAug 24, 2026

More Motion Is Not Always Better Motion: Corpus Composition Governs Whether Augmentation Helps SMPL-Based Parkinsonian Gait Severity Estimation

Authors: Michael Caiola, Andrew C. Weitz

Organizations: Credence, McLean VA, USA

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: Inference path. A walk’s SMPL sequence is rendered to 2D joint projections from two synthetic camera views, cut into 81-frame clips, and embedded by three frozen MotionAGFormer encoders. Embeddings are averaged over clips and views, standardized, and passed to per-member MLP heads whose softmax outputs are averaged. The encoders share one WearGait-augmented lifting corpus and differ only in backbone size and seed.

	macro-F1	class-0	class-0		hidden
Lifting encoder	(0/1/2)	F1	recall	acc.	test
liftbase (no augmentation)	0.374 $\pm$ 0.005	0.495	0.577	0.410	0.51
liftbedlam (+ synthetic)	0.363 $\pm$ 0.006	0.401	0.459	0.378	0.43
liftwg (+ WearGait)	0.393 $\pm$ 0.001	0.510	0.640	0.422	0.53
liftwglong (+ WearGait, long)	0.406 $\pm$ 0.006	0.528	0.658	0.419	0.47

Table 1: Lifting-corpus ablation. Offline columns: LOCO over the four labeled CARE-PD cohorts (2950 walks), three seeds, one ensemble member; macro-F1 over classes 0/1/2 because class 3 appears in only two labeled cohorts. Last column: four-class macro-F1 on the hidden test, one evaluation of the encoder alone. Only the lifting corpus differs between rows.

Pool	Subjects	Sequences	Frames	Hurried-pace	Whole trials	Hidden-test
				share	or clips	macro-F1
liftspseg	119	3012	0.45 M	0%	clips	0.32
liftwgsp	119	692	0.85 M	0%	whole	0.35
liftwgmore	120	928	1.31 M	25.4%	whole	0.40
liftmixseg	120	3246	0.52 M	33.8%	clips	0.46
liftwgall	155	1716	2.66 M	27.5%	whole	0.47

Table 2: WearGait pool composition against the hidden test. The six top rows share the same recordings, sensors and inertial-to-SMPL conversion, differing in walking tasks and in whole trials versus short clips. Scores are hidden-test macro-F1, one evaluation per pool, each as a single encoder (not comparable with ensemble figures elsewhere), against 0.51 unaugmented. liftwg is the pool in the final system. The last row adds a third collection site at fixed task composition;

\dagger

its score is the three-member ensemble, as the single encoder was not evaluated.

Figure 2: The six WearGait pools of Table 2 against the hidden test. Left: hidden-test macro-F1 against the pool’s share of hurried-pace walking; among the four whole-trial pools, more hurried-pace walking means a higher score. Right: the same six pools against corpus size, which shows no such pattern; the largest pool stays below the no-augmentation baseline. One evaluation per pool, each as a single encoder.

Small member	macro-F1 (0/1/2)	class-0 F1	class-0 recall
liftbase	0.391 $\pm$ 0.011	0.496	0.599
liftbedlam	0.362 $\pm$ 0.019	0.430	0.490
liftwg	0.365 $\pm$ 0.014	0.465	0.609
liftwglong	0.372 $\pm$ 0.012	0.486	0.621
no lifting member	0.357 $\pm$ 0.020	0.411	0.519

Table 3: Lifting encoder inside the three-member ensemble. LOCO over the four labeled CARE-PD cohorts, three seeds, macro-F1 over classes 0/1/2. The two large members and the rest of the pipeline are held fixed, and only the small member varies. Offline only: not every small member was evaluated on the hidden test inside this ensemble.

	composition	hidden-test macro-F1
single	base ( $16\times 128$ )	0.51
single	small ( $26\times 64$ , liftwg )	0.53
pairs without the small encoder	base $+$ large	0.51
pairs without the small encoder	large $+$ xlarge	0.52
pairs with the small encoder	small $+$ base	0.54
pairs with the small encoder	small $+$ large	0.56

Table 4: Ensemble composition against the hidden test. Learned-feature members are lifting-trained on the same WearGait-augmented corpus and differ in backbone size and random seed; one evaluation per composition. Sizes are layers

\times

width; xlarge is 40 layers. The last block replaces the learned features with hand-crafted descriptors.

extent of change	variant	offline gain	hidden test
none	frozen encoders	—	0.58
features recentered	per-site, forced	—	0.52
	pooled over the test set	—	0.49–0.51
input geometry	six camera views	$+0.044$	0.52
top 2 of 26 blocks	partial unfreeze	$+0.030$	0.52
lifting corpus	three-site, 178 subjects	$-0.014$	0.50

Table 5: How much of the frozen representation each variant changed, against its hidden-test score. Offline gains are LOCO over the four labeled CARE-PD cohorts, three seeds, macro-F1 over classes 0/1/2. Hidden-test scores are for the full three-member ensemble with the change applied, so 0.58 is the unmodified final system.

Figure 3: Where external motion sources land in the final system’s predicted-grade space; the dashed lines are its mean predicted grade on CARE-PD walks of classes 0–3 (

0.44

1.01

1.61

2.96

). Each source appears under both recovery methods, joined by a line. Every healthy external source sits about a grade above the class-0 reference, and the spread between recovery methods within a source (

0.05

0.12

) is a fraction of that offset. The recorded walker walks land at the class-3 reference, matching the assistive-device criterion from kinematics alone. BEDLAM carries the organizers’ own ground-truth SMPL and has no recovery step.

Figure 4: Offline movement against hidden-test movement for the six experiments measured under both protocols. Circles are single lifting encoders against the unaugmented baseline (offline: LOCO macro-F1 over classes 0/1/2; hidden test: single-encoder evaluations against 0.51). Diamonds are changes to the final ensemble (offline gains as in Table 5 ; hidden test against 0.58), so deltas are comparable within a marker family rather than across the two. Offline selection would have picked the two largest offline gains; both lost.

Synthetic Data Augmentation

Synthetic data augmentation creates or modifies training examples to improve learning when real data are scarce, imbalanced, costly, or sensitive.

See all 454 papers

3rd Place Solution to Human Motion Challenges in Real-World and Clinical Settings (MoCha) @ECCV2026: Language-Aligned Motion Representations for Domain-Generalizable UPDRS-Gait Severity Estimation
Sep 9, 2026Soojie Kim, Muhammad Munsif, Minkyung Kim +1Representation Learning Medical Diagnosis
UNIST, South Korea
Aggregate, Don't Adapt: Subject-Level Posterior Aggregation and Transductive Calibration for Cross-Site Parkinsonian Gait Severity
Aug 20, 2026Junlong ShenGait Analysis Parkinson's Disease
University of Alberta, Edmonton, Canada
LLM-Conditioned Synthesis of Pathological Gaits via Structured Gait-Language Representations
Jun 4, 2026Mritula Chandrasekaran, Sanket Kachole, Jarek Francik +1Synthetic Data Augmentation Synthetic Data Generation
School of Computer Science and Mathematics, Kingston University, London, United Kingdom

Sep 9, 2026J/K move · Enter open · S save

3rd Place Solution to Human Motion Challenges in Real-World and Clinical Settings (MoCha) @ECCV2026: Language-Aligned Motion Representations for Domain-Generalizable UPDRS-Gait Severity Estimation

Soojie Kim, Muhammad Munsif, Minkyung Kim, Seungryul Baek

UNIST, South Korea

Representation Learning Medical Diagnosis Domain Generalization Gait Analysis Parkinson's Disease

Open paper AskPDF arXiv

In this work, we introduce language-aligned motion representations for domain-generalizable UPDRS-Gait severity estimation, aiming to learn semantically structured motion features that generalize across heterogeneous clinical domains. We first learn motion representations using a Bi-GRU backbone that captures the temporal dynamics of SMPL sequences. Prior to model training, motion captions are generated offline using Qwen2.5-7B-Instruct. The backbone is then trained with both classification and text-alignment objectives to learn discriminative and semantically structured motion representations while accounting for the class imbalance present in the training data. We subsequently adapt the learned backbone independently to each source domain so that the model can capture domain-specific motion characteristics. The resulting source-specific models are then merged at the parameter level to consolidate complementary knowledge across source domains into a single domain-generalized model. To further mitigate class imbalance, we perform GPT-5.5-based pseudo labeling, and our final merged models for each site do not use any class-prior correction during inference. The resulting model is evaluated under the unseen-site setting of the MoCha Challenge, using Macro F1 as the primary evaluation metric. Our method achieves a macro-F1 of 0.57 on the hidden test set with only 637K active parameters at inference, ranking 3rd among 58 leaderboard entries in the MoCha 2026 Challenge. The challenge attracted 1,669 submissions from 112 participants and offered monetary prizes sponsored by Machine Medicine Technologies.

Abstract

Synthetic Data Augmentation

Explore similar work

3rd Place Solution to Human Motion Challenges in Real-World and Clinical Settings (MoCha) @ECCV2026: Language-Aligned Motion Representations for Domain-Generalizable UPDRS-Gait Severity Estimation

Aggregate, Don't Adapt: Subject-Level Posterior Aggregation and Transductive Calibration for Cross-Site Parkinsonian Gait Severity

LLM-Conditioned Synthesis of Pathological Gaits via Structured Gait-Language Representations