Decoding Children's Gait Behavior
Yifan Shen, Boyi Li, Meihuan Huang, Yuanzhe Liu, Xu Cao, Jinyang Jin, Zhengyuan Li, Anglin Liu, Junho Kim, Jingyuan Zhu, Lan Fangzhou, Jianguo Cao, Jintai Chen, Ismini Lourentzou, James Matthew Rehg
ECCV 2026
cs.CV
2026-08-01
Releases CGV, a 110-subject 60FPS pediatric gait dataset, with ChildGait-Video reaching 84% accuracy on EVGS scoring while frontier VLMs like GPT-5.2 and Gemini 3 sit near chance.
Gait analysis matters for diagnosing and treating developmental and neuromuscular disorders in children; cerebral palsy and hemiplegia both rely on it. Current systems need 3D sensors, which are expensive, intrusive, and impractical for young children. Computer-vision gait models are almost all pretrained on adult data and built for mature, stable, highly periodic adult gait. Children's gait has higher entropy, larger intra-class variance, and unstable patterns, so these models transfer poorly.
This paper frames fine-grained pediatric gait analysis from ordinary RGB video as a new vision problem domain, with the Edinburgh Visual Gait Score (EVGS) as the clinical target, 17 items per limb. To support the work it releases the Children Gait Video (CGV) dataset.
ChildGait-Video is an end-to-end model built on VideoMAE v2 that maps visual features directly to EVGS items. Two design choices do the work:
The authors deliberately avoid the geometric pipeline (extract pose, compute geometry, score by rules) because it discards appearance and texture and evaluates joint angles in isolation, missing kinematic synergy between joints. An MLP head predicts the EVGS items from the visual features.
The CGV dataset: 110 pediatric patients (ages 2.6 to 16.6, mean 8.4), 1,185 videos, 339,236 frames, 1920x1080 at 60 FPS, sagittal and coronal views, a 5-second walk-around task per session. 17 EVGS items per limb, 34 per patient. Diagnoses span cerebral palsy, traumatic brain injury, hip dysplasia, toe walking. Because of class imbalance, moderate and severe are merged into a normal-vs-atypical binary.
| Method | Left limb 17-item accuracy | Right limb 17-item accuracy |
| GaitSet / GaitPart / GaitGL / GaitBase | 48% to 51% | n/a |
| BiggerGait (prior SOTA gait model) | 54% | 53% |
| VideoMAE v2 | 69% | 72% |
| ChildGait-Video | 84% | 84% |
ChildGait-Video reaches F1 of 0.83, with per-item accuracy between 70% and 93% across the 34 items. A McNemar test against VideoMAE v2 gives p = 2.3e-4, a statistically significant gain. The human expert baseline is 93.8% accuracy with ICC = 0.93.
The frontier VLM numbers are more telling: zero-shot, Gemini 3 Pro scores 53%/55%, GPT-5.2 53%/56%, Qwen3-VL-235B 54%/59%, all sitting near the chance line for a binary task; a fine-tuned Qwen3-VL-ChildGait adds only 1 to 2 points. A frame-count ablation gives 74.1% at 8 frames (0.26s), 87.0% at 16, and 90.7% at 32, showing clinical gait assessment needs enough temporal context.
The main contribution is problem definition plus dataset. It documents an overlooked fact with measured numbers: even frontier multimodal LLMs are close to guessing when asked to read fine-grained biomechanical deviations from video, because they lean on semantic understanding and are weak at quantitative biomechanical interpretation. Specialized gait recognition networks, built for identity, are no better. An end-to-end video model with simple priors beats them by a wide margin. The clinical payoff would be replacing expensive 3D gait systems with a single ordinary camera for pediatric screening.
Younger children are harder: the under-8 group reaches 82.4% versus 85.2% for 8 and up, so still-developing motor patterns are tougher to model. The work is model-level validation only, with data collected in hospitals; generalization to open scenes like homes, rehab centers, and communities is not yet tested. The domain gap from adult pretraining persists. More fundamentally, the paper exposes the VLM weakness on fine-grained action recognition without solving it.