FoundationGait: Scalable Gait Model

• The paper introduces FoundationGait as the first scalable, self-supervised gait foundation model that uses part-aware training to preserve local body-part diversity.
• It employs a teacher–student pretraining framework with a mixed-domain corpus, achieving significant improvements in zero-shot and fine-tuned performance across recognition and healthcare tasks.
• The framework overcomes prior anti-scaling issues by maintaining structured local variation, which enhances transferability and performance across various gait applications.

Searching arXiv for papers on FoundationGait and related gait foundation models. FoundationGait is a gait foundation model introduced as the first explicitly scalable, self-supervised framework for silhouette-based gait understanding. It is designed to address two barriers that had limited prior gait systems: scalability and generalization. In the formulation associated with the model, gait analysis had not failed to follow a foundation-model trajectory because gait is intrinsically too simple, but because prior architectures and training schemes did not preserve the fine-grained local body-part structure on which gait discrimination depends. FoundationGait therefore combines a mixed-domain pretraining corpus, a teacher–student self-supervised objective, and a part-aware training strategy intended to preserve local feature diversity as model size increases, while supporting transfer across identification and healthcare tasks such as gait recognition, scoliosis screening, depression prediction, and gait attribute estimation (Ye et al., 30 Nov 2025).

1. Origins and problem formulation

FoundationGait was proposed against a background in which gait research had largely split into isolated subfields. Identification systems emphasized rank-based retrieval under viewpoint, clothing, and carrying variation, whereas healthcare-oriented systems emphasized disorder-specific supervision, often on small and narrow datasets. The FoundationGait formulation treats that separation as artificial: both settings depend on subtle motion and shape regularities in walking, and a single pretrained model should ideally transfer across datasets, conditions, tasks, and even input modality (Ye et al., 30 Nov 2025).

A central claim of the framework is that naïve model scaling had historically failed in gait. The largest prior gait models were usually under 5M parameters, and simple enlargement of representative supervised and self-supervised baselines did not yield monotonic gains. A common misconception is that low-resolution binary silhouettes are too simple to benefit from scale. FoundationGait argues instead that the main failure mode is feature homogenization: deeper or wider gait networks amplify only the most discriminative early features, suppress weaker but important local cues, and reduce diversity among part representations. This diagnosis is significant because it reframes gait foundation modeling as a problem of preserving structured local variation, not merely adding parameters (Ye et al., 30 Nov 2025).

In this sense, FoundationGait is not defined only by dataset size or parameter count. Its stated objective is to place gait understanding on a scaling-and-transfer trajectory by coupling large-scale self-supervised pretraining with architectural mechanisms that remain sensitive to body-part structure. A plausible implication is that the model is as much a proposal about inductive bias as it is about scale.

2. Architecture and self-supervised pretraining

FoundationGait uses the DeepGaitV2 backbone as its encoder and scales it by increasing layer depth. Inputs are silhouettes normalized to $64 \times 44$, processed as clips of 16 frames during pretraining. The pretraining setup is a teacher–student framework with an encoder $F$, projector $H$, and predictor $G$ on both branches. Given two non-overlapping clips $c_i$ and $c_{i+}$ sampled from the same sequence and augmented by $\Pi(\cdot)$, the teacher branch produces

$$k_{i+} = H_t(\mathrm{Pool}(F_t(\Pi(c_{i+})))), \qquad q_{i+} = G_t(k_{i+}),$$

while the student branch first applies part-aware splitting:

$$\{v_{i,j}\}_{j=1}^{p} = F_s(\mathrm{Split}(\Pi(c_i))), \qquad v_i = \mathrm{Concat}(\{v_{i,j}\}_{j=1}^{p}),$$

and then computes

$q_i = G_s(H_s(\mathrm{Pool}(v_i))).$

The pretraining objective is an InfoNCE contrastive loss between the teacher representation and the student predictor output:

$F$0

with cosine similarity and $F$1 (Ye et al., 30 Nov 2025).

The distinctive mechanism is the part-aware student. Rather than feeding the second view directly through a conventional student encoder, the model horizontally splits the silhouette clip into $F$2 parts, processes those parts independently, and concatenates them back into a global feature map. During pretraining, FoundationGait uses multiple granularities $F$3, assigning different sub-batches to different $F$4 values, and applies a shifted-window-like operation to half the samples during splitting. This design forces the student to build a global representation from detail-enriched local feature maps. The intended effect is to preserve part diversity even as the model becomes larger, thereby reversing the anti-scaling behavior observed in earlier gait architectures (Ye et al., 30 Nov 2025).

Four model scales are reported. The 0.006B variant has 38.5M total parameters, the 0.01B variant 43.7M, the 0.03B variant 65.7M, and the largest 0.13B variant 164.9M total parameters with 132.3M backbone parameters and 24.21G FLOPs on $F$5 silhouette input. The abstract describes the largest model as “nearly 0.13 billion” parameters, while the detailed configuration reports 164.9M total parameters (Ye et al., 30 Nov 2025).

3. Pretraining corpus and scaling behavior

FoundationGait is pretrained on WebGait-2M, a mixed corpus built from 12 public train sets spanning recognition and healthcare tasks. The datasets are CASIA-B, CCPG, CASIA-E, OU-MVLP, SUSTech1K, CCGR, Gait3D, GREW, D-Gait, RA-GAR, Scoliosis1K, and GaitLU-1M. Altogether, WebGait-2M contains 2,358,547 sequences and 231,675,844 frames. The pretraining procedure treats each sequence as an individual ID in the unsupervised setting, samples datasets with probabilities proportional to $F$6, and downweights GaitLU-1M by $F$7 because of its scale. The 0.13B model is trained with batch size 512, 16-frame clips, SGD with learning rate 0.05, momentum 0.9, weight decay $F$8, and cosine annealing for 80K iterations, requiring about 178 hours on $F$9 32GB V100 GPUs (Ye et al., 30 Nov 2025).

The scaling analysis is one of the framework’s main scientific claims. The authors show negative evidence first: enlarged versions of DeepGaitV2 and GaitSSB do not improve monotonically and can degrade. In zero-shot transfer on WebGait-2M, GaitSSB reaches 44.4% rank-1 on OU-MVLP, whereas GaitSSB-0.03B drops to 39.2% and GaitSSB-0.13B to 38.9%. On CASIA-B CL, GaitSSB obtains 33.5%, but larger variants fall to 26.0% and 25.9%. FoundationGait is presented as reversing that trend. Moving from FoundationGait-0.03B to FoundationGait-0.13B, zero-shot OU-MVLP rank-1 rises from 57.0% to 64.5%, Gait3D rank-1 from 41.1% to 48.0%, CASIA-B NM from 92.0% to 94.0%, and CCPG DN from 61.8% to 64.9%. A plausible implication is that the framework’s principal contribution is not merely large-scale pretraining, but a specific strategy for making scaling work in a part-based gait regime (Ye et al., 30 Nov 2025).

The model is therefore framed as evidence that gait models can follow an empirical scaling law, even though the paper does not fit an explicit power-law regression. The supporting evidence is monotonic improvement with scale under the part-aware design, contrasted with anti-scaling behavior in naïvely enlarged baselines.

4. Transfer across recognition, healthcare, and modality

FoundationGait is evaluated in zero-shot, fine-tuned, healthcare-transfer, and cross-modality settings. In zero-shot recognition, the pretrained model is used without supervised adaptation on CASIA-B, OU-MVLP, Gait3D, GREW, CCPG, and SUSTech1K. The headline results are 48.0% zero-shot rank-1 on Gait3D with 1,000 test subjects and 64.5% rank-1 on OU-MVLP with over 5,000 test subjects. The same 0.13B model also reaches 66.5% Gait3D rank-5, 29.4% GREW rank-1, 49.0% SUSTech1K rank-1, and 27.6/39.2/64.9/80.8% on CCPG under CL/UP/DN/BG conditions (Ye et al., 30 Nov 2025).

For downstream recognition fine-tuning, the teacher branch is used as initialization, the predictor is replaced with a task head, and part-aware training remains active. The recognition objective combines triplet loss and cross-entropy:

$H$0

$H$1

Under this regime, FoundationGait-0.13B reaches 87.2% mean accuracy on CCPG, 79.3% rank-1, 89.5% rank-5, and 74.0% mAP on Gait3D, and 53.25% rank-1, 48.00% mAP, and 33.29% mINP on CCGR-MINI (Ye et al., 30 Nov 2025).

The healthcare transfer experiments justify the broader “foundation model” designation. On RA-GAR, the 0.13B linear probe achieves 85.7% instance accuracy, 79.3% precision, 77.3% recall, 78.2% F1, and 65.6 mA. On D-Gait, linear probing gives 57.8% F1, and fine-tuned FoundationGait-0.03B reaches 75.1% accuracy, 67.9% precision, 65.1% recall, and 66.1% F1. On Scoliosis1K, fine-tuned FoundationGait-0.03B reaches 97.3% accuracy, 95.4% precision, 88.9% recall, and 91.7% F1, outperforming the cited prior baseline ScoNet-MT on both accuracy and F1. The paper also notes a scale-related class-imbalance failure mode: the larger 0.13B model can become overly optimistic about positive predictions on imbalanced healthcare tasks, leading to high recall but reduced precision and F1 (Ye et al., 30 Nov 2025).

Cross-modality transfer is tested by fine-tuning the silhouette-pretrained model on human parsing inputs. Replacing the input with a $H$2 parsing map while keeping the training setup otherwise unchanged yields 83.8% rank-1, 92.7% rank-5, and 79.3% mAP on Gait3D for the parsing-only FoundationGait-0.13B. A simple test-time fusion of parsing and silhouette models by averaging metric distances reaches 86.5% rank-1 and 82.6% mAP, exceeding earlier multimodal baselines reported in the paper (Ye et al., 30 Nov 2025).

5. Relation to adjacent gait modeling work

FoundationGait emerges from a broader trajectory in which gait representations became progressively more structured, more generalizable, and less tied to single datasets or modalities. Several earlier works supply component-level ideas that clarify what FoundationGait adopts and what it departs from.

“GaitFormer” emphasizes the intrinsic periodicity of gait and introduces Temporal Periodic Alignment, including Adaptive Fourier-transform Position Encoding and a Temporal Aggregation Module. Its central point is that gait should not be treated as an arbitrary sequence problem, but as one organized by repeated motion cycles (Wu et al., 2023). “Motion Gait” argues for explicit motion-change modeling through a parameter-free Motion Excitation Module and part-specific local representation learning through a Fine Feature Extractor, particularly under clothing and carrying changes (Zhang et al., 2022). “GPGait” focuses on cross-dataset generalization in pose-based recognition through Human-Oriented Transformation, Human-Oriented Descriptors, and a Part-Aware Graph Convolutional Network, identifying camera-coordinate dependence as a major source of transfer failure (Fu et al., 2023). “On Denoising Walking Videos for Gait Recognition” introduces diffusion-based denoising and a geometry-driven Gait Feature Field to remove identity-irrelevant RGB cues while retaining motion-relevant structure (Jin et al., 24 May 2025).

These works suggest a set of inductive biases that FoundationGait elevates into a large-scale self-supervised setting: periodicity, motion saliency, part-aware locality, human-centered canonicalization, and nuisance suppression. A plausible implication is that FoundationGait should be read not as a wholly discontinuous break with prior gait modeling, but as an overview that scales several previously separate design principles (Ye et al., 30 Nov 2025).

The broader field also shows that “gait foundation model” can denote more than silhouette-based identification transfer. A later paper on 3D skeletal motion develops a self-supervised gait foundation model from 3,414 deeply phenotyped adults and shows that gait embeddings predict age, BMI, visceral adipose tissue area, thousands of phenotypic targets, and medication use, with anatomically structured attribution across legs and torso (Gabet et al., 26 Mar 2026). This suggests that FoundationGait’s transfer across healthcare tasks is part of a larger shift in which gait is treated as a reusable systems-level biosignal rather than a narrow biometric or pathology label.

6. Limitations, open problems, and interpretation

FoundationGait is relatively explicit about its limitations. The authors identify two open problems directly: a ViT-based gait foundation model remains unresolved, and scaling beyond 0.13B toward the 1B regime is left as ongoing work. An ablation replacing the last two CNN blocks with a 12-layer ViT hurts performance substantially, which the paper interprets as evidence that the current success of the framework depends in part on CNN-style local inductive bias rather than on scale alone (Ye et al., 30 Nov 2025).

The framework is also silhouette-based. This gives strong control over nuisance appearance information, but it leaves open how a gait foundation model should integrate richer modalities such as RGB, pose, structural vibration, wearable IMU, plantar pressure, or force signals. Other work in the supplied literature points toward that need. Structural-vibration sensing has been used to infer foot-floor contact type and duration from floor response spectra, offering a contact-centric intermediate representation with force-informed content (Dong et al., 2024). Remote pathological gait classification from GEI and SEI templates shows the value of lightweight deployment but remains limited by simulated pathology and studio data (Albuquerque et al., 2021). GAVD provides clinically annotated in-the-wild RGB video of abnormal gait, while GAIT-IT provides cleaner supervised pathology templates, suggesting complementary but still limited sources of clinical supervision (Ranjan et al., 2024, Albuquerque et al., 2021).

Several omissions are also material. The FoundationGait paper does not provide a deeper fairness analysis across demographics, and although the supplementary material states that all experiments use public datasets collected with informed consent and that original licenses and privacy policies are respected, bias, fairness, and medical deployment risk remain largely open (Ye et al., 30 Nov 2025). In healthcare transfer, the model can over-predict the positive class under long-tailed label distributions. In recognition, its strongest results remain tied to standardized silhouette preprocessing and benchmark protocols rather than unconstrained multimodal sensing. These limitations do not negate the model’s status, but they delimit it.

FoundationGait is therefore best interpreted as a framework that establishes three propositions. First, gait models can benefit from scale when model design preserves local body-part diversity. Second, one pretrained gait model can transfer across identity and healthcare tasks. Third, self-supervised pretraining on a broad, mixed-domain gait corpus can produce directly useful zero-shot and fine-tuned representations. Its most important contribution is not only its reported zero-shot and downstream metrics, but the claim that gait understanding can be organized around scalable representation learning rather than isolated task-specific systems (Ye et al., 30 Nov 2025).