FootGait3D: Multi-View Foot-Ankle Dataset

• FootGait3D is a multi-view 3D point cloud dataset capturing detailed dynamic foot–ankle geometry during natural gait with high resolution and controlled occlusion.
• It employs a custom markerless five-camera depth sensing system to generate paired partial and complete point clouds for benchmarking 3D completion methods.
• The dataset supports both single-modal and multi-modal approaches under varied occlusion conditions, advancing research in biomechanics, gait analysis, and clinical applications.

FootGait3D is a multi-view dataset of high-resolution ankle-foot surface point clouds captured during natural gait, designed to provide paired partial and complete point clouds under realistic occlusion and to serve as a benchmark for 3D point cloud completion methods and for biomechanics and gait research (Li et al., 15 Jul 2025). It focuses specifically on the detailed modeling of the ankle-foot region rather than whole-body or generic lower-limb representations, and consists of 8,403 point cloud frames collected from 46 subjects using a custom five-camera depth sensing system (Li et al., 15 Jul 2025). Within the recent gait-analysis literature, it occupies a distinct position between whole-body 3D gait datasets built around SMPL or silhouettes, such as Gait3D with 4,000 subjects and over 25,000 sequences from 39 cameras (Zheng et al., 2022), and dynamic foot-specific 4D datasets such as “4D Feet,” which includes 4D shapes at 15 fps of the right and left feet of 58 participants and 5,147 3D frames (Tajdari et al., 2023).

1. Definition, scope, and positioning

FootGait3D is defined as a real-world, multi-view 3D point cloud dataset of the human foot–ankle complex during gait, with an explicit emphasis on paired partial and complete point clouds under realistic occlusion (Li et al., 15 Jul 2025). Its target applications are stated as 3D point cloud completion under occlusion, biomechanics and multi-segment foot modeling, clinical gait analysis, prosthetic and orthotic design, robotics and humanoid locomotion, and markerless motion capture research (Li et al., 15 Jul 2025). The dataset is publicly released at https://huggingface.co/datasets/ljw285/FootGait3D (Li et al., 15 Jul 2025).

The scope is anatomically distal and geometrically dense. Unlike datasets centered on silhouettes, sparse skeletons, or whole-body SMPL parameters, FootGait3D focuses exclusively on the foot–ankle complex and captures detailed surface geometry, including toes, heel, malleoli, arch, and plantar surface (Li et al., 15 Jul 2025). This focus differentiates it from markerless monocular gait systems that estimate bespoke gait parameters such as inversion/eversion, dorsiflexion/plantarflexion, ankle angle, and foot progression angle from a single RGB camera (Wang, 2020). A plausible implication is that FootGait3D is better suited to problems in which the distal surface itself is the primary object of study, rather than a latent input to joint-angle estimation.

The authors position FootGait3D as a bridge between computer vision completion research and biomechanics. In that sense it complements, rather than replaces, whole-body 3D gait recognition frameworks such as SMPLGait, which exploit dense 3D body representations for identity recognition in unconstrained scenes (Zheng et al., 2022). It also complements foot-specific reconstruction work on incomplete self-scans, where canonicalization and learned completion are applied to static or mobile-capture foot geometry (Fogarty et al., 27 Feb 2025). FootGait3D instead contributes dynamic, stance-phase, multi-view, real sensor data with controlled view dropout.

2. Acquisition system and geometric formulation

FootGait3D is collected with the Point-cloud Foot Analysis system, a custom markerless setup using five Occipital Structure Core depth sensors (Li et al., 15 Jul 2025). The sensors have a field of view of $59^\circ \times 46^\circ \times 70^\circ$, a capture rate of 30 FPS, depth precision of about 3 mm within 1 m, and reconstruction RMSE below 2 mm (Li et al., 15 Jul 2025). Four horizontal sensors, DS0–DS3, are placed around the capture zone and aimed at the dorsal and side aspects of the foot, while one bottom sensor, DS4, is mounted under a transparent gait plate and tilted upwards to capture the plantar surface during stance (Li et al., 15 Jul 2025).

All point clouds are transformed into a shared global coordinate system whose origin is the geometric center of the upper surface of the transparent plate; the axes are $X$ for anterior–posterior, $Y$ for superior–inferior, and $Z$ for medial–lateral (Li et al., 15 Jul 2025). Calibration follows a coarse-to-fine registration procedure using a custom 3D calibration object for coarse registration and point-to-plane ICP for fine registration. For a sensor point cloud $P=\{p_i\}$ and reference cloud $Q=\{q_j\}$ with normals $n_j$, the fine registration minimizes

$$E(R, t) = \sum_i \left( \big( R p_i + t - q_{\text{NN}(i)} \big) \cdot n_{\text{NN}(i)} \right)^2.$$

Depth-to-3D conversion uses the camera intrinsics $K$ and the extrinsic transform $T_{\text{cam}\to\text{world}}$. For a depth pixel $X$0 with depth $X$1,

$X$2

These definitions matter because FootGait3D is not a mesh-parameter dataset or a pose-estimation dataset; it is a calibrated multi-view surface dataset. A plausible implication is that it supports downstream tasks requiring explicit spatial consistency across views, including registration, non-rigid tracking, and physical model fitting.

3. Gait protocol, preprocessing, and dataset composition

The dataset contains 46 healthy adults with mean age $X$3 years, mean height $X$4 cm, and mean weight $X$5 kg; inclusion required no lower-limb, foot, or ankle injuries or diseases, and all participants provided written informed consent under IRB approval from Fudan University (Li et al., 15 Jul 2025). Participants walked barefoot, overground, back and forth along the walkway at self-selected comfortable speed, and were instructed to ensure the foot made full contact with the transparent plate inside the capture zone (Li et al., 15 Jul 2025). Each subject completed 8 gait tests after familiarization, and each stance-phase sequence typically contains 28–36 frames at 30 FPS (Li et al., 15 Jul 2025).

The preprocessing pipeline converts each saved depth map into a single-view point cloud, groups frames across cameras by software timestamps, fuses the transformed clouds in the global frame, applies spatial cropping with a fixed 3D bounding box, detects stance phase by counting foot points in a ground layer, and removes the swing foot by thresholding in the medial–lateral direction (Li et al., 15 Jul 2025). Heel strike and toe-off are identified by sudden changes in the number of points within the ground layer, and only frames between heel strike and toe-off are kept (Li et al., 15 Jul 2025). Manual inspection and additional cropping are then used for quality control (Li et al., 15 Jul 2025).

For each retained time instance, the authors define a full-view ground-truth cloud as the fusion of DS0–DS4. Partial point clouds are created by removing DS4 and fusing subsets of DS0–DS3: one 4-view case, four 3-view cases, and six 2-view cases (Li et al., 15 Jul 2025). This yields graded occlusion scenarios in which the plantar surface is systematically absent from the partial inputs.

Component	Quantity	Notes
Subjects	46	Healthy adults
Full-view stance frames	8,403	5-view ground truth
Total point cloud frames	100,836	5-view, 4-view, 3-view, 2-view combined

The geometry density is unusually high for a gait dataset. The 5-view ground truth clouds mostly contain about 78,000–120,000 points, 4-view clouds about 60,000–106,000 points, 3-view clouds about 40,000–80,000 points, and 2-view clouds about 20,000–60,000 points, with more variability and sparser outliers in the 2-view condition (Li et al., 15 Jul 2025). The released split is instance-based rather than explicitly subject-based: 5,881 training instances, 841 validation instances, and 1,681 test instances (Li et al., 15 Jul 2025). The authors note that strict subject-wise generalization would require a re-split by subject (Li et al., 15 Jul 2025).

4. Completion benchmark and evaluation protocol

The core task in FootGait3D is 3D point cloud completion under occlusion (Li et al., 15 Jul 2025). The input is a partial point cloud generated from a 4-view, 3-view, or 2-view fusion, and in the benchmark these partial clouds are downsampled to 2,048 points by uniform sampling or Farthest Point Sampling (Li et al., 15 Jul 2025). The output is a completed point cloud approximating the 5-view ground truth, typically with either 16,384 points or 2,048 points depending on the network family (Li et al., 15 Jul 2025).

Evaluation uses Chamfer Distance and an $X$6-score at threshold $X$7, where $X$8 is defined as $X$9 of the bounding box diagonal in normalized units (Li et al., 15 Jul 2025). For predicted set $Y$0 and ground-truth set $Y$1, the reported metrics are

$Y$2

$Y$3

and

$Y$4

When models produce different output sizes, both prediction and ground truth are downsampled to 2,048 points via FPS before computing CD and $Y$5 (Li et al., 15 Jul 2025). This is important because FootGait3D compares single-modal and multi-modal completion networks with different decoder resolutions, and the evaluation protocol normalizes that discrepancy at test time.

The benchmark formulation also supports adjacent tasks, including denoising, registration and tracking, gait-phase characterization, and model fitting (Li et al., 15 Jul 2025). This suggests that FootGait3D can function not only as a completion benchmark but also as a calibrated substrate for non-rigid correspondence and dynamic shape modeling.

5. Baseline models and empirical behavior

The benchmark includes two categories of completion models (Li et al., 15 Jul 2025). The single-modal, geometry-only group comprises PCN, PoinTr, AnchorFormer, SnowflakeNet, and PointAttn; all take only the partial point cloud as input (Li et al., 15 Jul 2025). The multi-modal group comprises SVDFormer, PointSea, CSDN, EGIINet, XMFNet, and MAENet; these use the partial point cloud together with multi-view depth maps and, where applicable, real extrinsic matrices from the five-camera setup (Li et al., 15 Jul 2025). The authors state that all methods are trained with their official hyperparameters and loss formulations (Li et al., 15 Jul 2025).

The reported quantitative pattern is consistent across occlusion levels. Among single-modal methods, PCN is the weakest, while PointAttn and SnowflakeNet are stronger; PointAttn achieves $Y$6, $Y$7, and $Y$8 in the 3-view setting, and similar performance in the 2-view setting (Li et al., 15 Jul 2025). Among multi-modal methods, SVDFormer and PointSea are the strongest and most stable. PointSea reports $Y$9, $Z$0, and $Z$1 in the 4-view setting, with only minor degradation in the 3-view and 2-view settings (Li et al., 15 Jul 2025). SVDFormer shows a closely similar pattern, including $Z$2, $Z$3, and $Z$4 in the 4-view setting (Li et al., 15 Jul 2025).

Several qualitative findings are central. First, all methods perform worse on FootGait3D than on PCN or ShapeNet, which the authors interpret as a real-versus-synthetic gap for deformable, noisy anatomical data (Li et al., 15 Jul 2025). Second, multi-modal methods consistently outperform single-modal baselines, confirming the benefit of multi-view depth information (Li et al., 15 Jul 2025). Third, robustness to reduced views varies by architecture: PointSea and SVDFormer remain stable across 4-view, 3-view, and 2-view inputs, whereas AnchorFormer is more sensitive to severe sparsity (Li et al., 15 Jul 2025).

A plausible interpretation is that FootGait3D stresses not only geometric completion capacity but also the ability to exploit sensor geometry, real view configurations, and anatomically structured priors. That makes it a harder and more realistic testbed than generic rigid-object completion corpora.

6. Biomechanical relevance, limitations, and outlook

FootGait3D enables dynamic surface-level biomechanical analysis of the foot–ankle complex, including multi-segment foot models, dynamic morphology under load, footwear and orthotic design, and markerless clinical gait analysis (Li et al., 15 Jul 2025). The anatomically visible regions in the 5-view ground truth include the dorsal surface of toes and metatarsals, the lateral and medial malleoli, the calcaneus, the arch region, and the plantar surface from DS4 (Li et al., 15 Jul 2025). This makes the dataset relevant to applications that depend on distal geometry rather than solely on joint-center kinematics.

In relation to adjacent research, FootGait3D fills a gap between dense surface acquisition and clinically interpretable gait modeling. Work on personalized digital twins with soft-body feet has shown that subject-specific foot geometry can be integrated into walking simulation and evaluated with ground reaction force and joint angle results (Loke et al., 2024). Markerless gait-analysis frameworks based on 3D reconstruction and OpenSim have shown that biomechanically meaningful markers can improve agreement with marker-based gait measurements (Pemasiri et al., 3 Mar 2026). A plausible implication is that FootGait3D could support future pipelines in which dense foot–ankle surfaces are fitted to statistical or biomechanical models and then linked to kinematic or simulation frameworks.

The stated limitations are also precise. Cameras are synchronized in software only, so small temporal offsets may cause slight inconsistencies, especially at heel strike and toe-off; some gait cycles are incomplete because frames with obvious misalignment are manually removed (Li et al., 15 Jul 2025). The cohort contains 46 healthy young adults and does not include pathological gait, shod walking, or outdoor conditions (Li et al., 15 Jul 2025). The release does not include anatomical landmark labels, segment definitions, joint angles, or ground reaction forces (Li et al., 15 Jul 2025). These constraints distinguish FootGait3D from systems dedicated to direct gait-parameter estimation from RGB video (Wang, 2020) or from markerless 3D gait frameworks built for joint kinematics (Pemasiri et al., 3 Mar 2026).

The authors propose hardware-triggered synchronization, higher-frame-rate depth sensors, more complete gait cycles, more subjects, pathological conditions, shod conditions, various walking speeds, stairs or uneven terrain, and completion methods that better exploit temporal continuity and biomechanical priors (Li et al., 15 Jul 2025). More broadly, FootGait3D suggests a research direction in which foot completion, dynamic registration, and biomechanics are studied jointly. That direction is also consistent with recent foot-reconstruction work that uses learned geometric priors to recover clinically relevant anatomy from incomplete scans (Fogarty et al., 27 Feb 2025), but FootGait3D adds the dynamic, stance-phase, multi-view setting that such methods have generally lacked.