Fast dynamic scene renderers store each moving 3D point's path as a smooth curve over time. That design lets them answer "where is this point at any instant" without a neural network, and it is why they run at tens to hundreds of frames per second. Smoothness is the wrong prior for a foot, though. A planted foot sits exactly still for a stretch of time and then leaves abruptly. A smooth curve fitted through slightly jittery observations of that foot wobbles between frames, so the rendered foot glides across the floor ("foot sliding"). Because a still foot looks like background, the renderer may also absorb it into the static scene and later leave floating debris behind when it lifts.
We propose to make "stopped" a first-class state of the trajectory representation. Each foot-region trajectory becomes a sequence of free swing segments and exactly stationary stance segments pinned to a reconstructed floor plane. The switching times are learned from images and a contact detector. The claim is that one structural change in how time is represented can do three things at once. It removes foot sliding at arbitrary query times, not just at captured frames. It resolves the static/dynamic confusion at contact zones. It costs nothing in rendering speed, and may save some, because stance segments are constants.
Core Claim
The fully explicit dynamic Gaussian methods share a smooth-curve assumption. Here a 3D Gaussian is one of many small colored blobs rasterized to form an image. Fully Explicit Dynamic Gaussian Splatting (Ex4DGS) interpolates positions with cubic Hermite splines between keyframes. SplineGS gives each Gaussian a Hermite spline with adaptively pruned control points. AD-GS uses B-splines for rigid vehicles. All of them assume trajectories with continuous velocity and acceleration everywhere. None represents a finite interval of identically zero velocity, and the research file finds no work among them that adds contact-aware knots.
Contact methods do the reverse. Combining Recurrent Neural Networks and Adversarial Training penalizes foot velocity, and FECO predicts dense contact. Both operate only at discrete frames and know nothing about the radiance scene.
Our representation is a hybrid trajectory. For a foot-associated Gaussian or motion cluster $i$, the position $p_i(t)$ equals a fixed world anchor $q_{i,k}$ during stance interval $[a_k, b_k]$. Between stances it is a cubic Hermite segment whose end tangents are fixed to zero:
$$ p_i(t)=\begin{cases} q_{i,k}, & t\in[a_k,b_k] \\ H\big(q_{i,k},\,0,\;q_{i,k+1},\,0;\;t\big), & t\in(b_k,a_{k+1}) \end{cases} $$
Here $H(\cdot)$ is a Hermite segment from position $q_{i,k}$ with zero velocity at lift-off to position $q_{i,k+1}$ with zero velocity at touchdown, optionally with interior control points for the swing. This makes velocity exactly zero throughout stance and continuous at the switches. Acceleration jumps at the switches, which is the physically correct behavior at heel strike and toe-off. Anchors are constrained to lie on or above a support plane fitted to nearby static Gaussians.
What is new is not a zero-velocity loss. It is a representation in which sliding during stance is impossible by construction at every continuous time. The switching times are optimized jointly with appearance.
Research Program
The work has four parts.
Representation and switch learning. The switch times $a_k, b_k$ are first set from per-frame contact probabilities from a FECO-style dense foot contact estimator, smoothed over time. They are then refined by the photometric loss. During early training the hard case split is relaxed into a sigmoid blend between the stance constant and the swing segment, and its temperature is annealed toward a hard switch. A small sparsity term discourages spurious short stances.
Which Gaussians switch. Gaussians are attached to stance clusters in one of two ways. With an optional body model, they belong to foot vertices. Without one, they are grouped by contrastive motion identity in the style of Split4D, which learns object identity from consistent motion rather than 2D trackers. This keeps the method usable in general dynamic scenes and not only with skeleton-driven avatars.
Support plane. The plane is fitted by RANSAC over static Gaussians in a band around stance anchors. This borrows the planar-primitive reasoning of CRISP but keeps it inside the reconstruction. A one-sided penetration penalty acts on anchors and swing segments.
Contact-zone decomposition and densification. Ex4DGS decides whether a Gaussian is static or dynamic from instantaneous motion residuals. That rule misclassifies a planted foot. We replace it near contact with the long-horizon identity: a Gaussian in stance stays dynamic-but-stationary and is never merged into the background. Densification is forbidden from cloning foot-region error into the static set within the contact band. This directly targets the shared-Gaussian blurring that Split4D reports at shoe–floor interfaces.
Rendering cost. At render time, a stance Gaussian costs one lookup instead of a spline evaluation. We will measure whether temporarily treating stance clusters as static in the rasterizer's sort cache yields a speedup, and report it honestly if it does not.
Motion blur (optional extension). Because the trajectory is defined at every instant, it can be integrated over the shutter window. We can therefore render blurred training frames by averaging sharp renders along the trajectory, as BAD-Gaussians does for camera motion in static scenes. A contact-switched trajectory has a useful property here: a planted foot produces no blur, while a swinging foot does. That asymmetry is itself evidence for the switch times. This addresses the blur overfitting that SplineGS names as its main failure on in-the-wild video. It is secondary to the main claim and will be evaluated only after the core result holds.
Baselines
The naive reference is Ex4DGS run unmodified. It is the strongest fully explicit multi-view method in the file, reporting 32.11 dB PSNR and 62.1 FPS at 1352×1014 on N3V per Fully Explicit Dynamic Gaussian Splatting. For monocular sequences the reference is SplineGS. Both are the strongest competitors because they already have continuous-time explicit trajectories and real-time rendering. The only thing they lack is contact structure, so any gain is attributable to our change. We found no reported foot-sliding or contact metric on Lacuna for either method, so we will measure these ourselves. We run official code on identical frames and cameras, with the same iteration count and the same Gaussian budget cap.
The second baseline is discrete clamping. It runs Ex4DGS, detects contacts per frame with a FECO-style estimator, and snaps foot-region Gaussian keyframe positions to their stance mean before the spline is fitted. This isolates "contact at frames" from "contact in continuous time."
The third baseline is a body-anchored avatar. Gaussians are embedded on a body mesh as in FlashAvatar, which reaches 300 FPS for heads. The mesh is driven by a skeleton whose feet are refined by FootMR, which reports a 30.6% ankle angle error reduction on MOYO. Its foot joints are then contact-clamped per frame. This is the most serious competitor for human-only sequences, and it gets the same tuning budget as our method.
The fourth is a soft-penalty ablation. It keeps standard cubic splines and adds a contact-weighted velocity penalty evaluated at densely sampled continuous times. This isolates the structural switch from the mere presence of a contact signal.
We also run a fixed-switch ablation, with switch times frozen at detector output, which shows whether photometric refinement of timing matters.
Beating the baselines means three things together. First, a lower sliding metric during ground-truth stance, measured at held-out intermediate times. Second, contact F1 against ground truth above the per-frame detector the method was initialized from. Third, PSNR within 0.3 dB of unmodified Ex4DGS and no frame-rate loss. For reference, BSTRO reports contact F1 of 0.708 on RICH per Capturing and Inferring Dense Full-Body Human-Scene Contact, and FECO reports 0.577 on MMVP.
Why not just…?
Why not just clamp feet on the captured frames and keep the existing splines? Clamping fixes positions at keyframes, but a $C^2$ spline through clamped keyframes followed by an abrupt lift-off must still ring, because a cubic cannot be identically constant over an interval and then accelerate within one knot spacing. The discrete-clamping baseline, evaluated at held-out intermediate timestamps, measures exactly this residual sliding. If it already matches us at mid-frame times, our representation is unnecessary, and we will say so.
Why not just add a dense continuous-time velocity penalty to the existing spline? This is the cheapest real threat. With enough control points, a soft penalty could approximate a stance plateau. Ex4DGS and SplineGS both prune or space control points for compactness, though, and the soft-penalty ablation will show whether matching our sliding numbers requires more control points per foot Gaussian. It might also trade off swing accuracy. We report sliding against control-point count for both methods. If the penalty matches at equal memory, the contribution reduces to the decomposition and densification change, and we will reframe accordingly.
Why not just drive Gaussians from a cleaned body skeleton? For a single person with a reliable body fit, this may be competitive. Its foot clamp is still discrete, however, and it has nothing to say about the background floor, shoes that deviate from the body mesh, or the static Gaussians that Split4D shows blending at contact. We apply our switched trajectory to the skeleton's foot joints as an extra arm. If skeleton plus switched trajectory wins on human-only data, the time representation still stands as the contribution. The per-Gaussian variant then has to earn its place on scenes where body fitting is weak.
Why not just run a physics simulator? CRISP and LAGOON obtain physical plausibility through hours of reinforcement learning in Isaac Gym. CRISP does so over planar cuboids that discard appearance. Neither renders novel views, so neither competes on this question. We cite CRISP's reduction in simulator tracking failure from 55.2% to 6.9% only to show that contact-correct support geometry matters.
Proposed Experiments
The main testbed is multi-view human capture with visible floors: DNA-Rendering, SelfCap, and the Multi-Human sequences used by Split4D. For contact ground truth we use RICH, whose scans and contact labels let us score both sliding and contact F1. N3V is a no-regression check against the Ex4DGS 32.11 dB number. Monocular generality is tested on EMDB-style walking with SplineGS as the base.
The decisive check is a temporal super-resolution protocol. We train on every second or third frame and evaluate sliding, contact F1, and PSNR on the held-out intermediate frames. This is where per-frame methods should fail and a continuous stance representation should not.
The metrics are as follows. Sliding is the mean horizontal foot displacement per second during ground-truth stance. Penetration is the depth below the fitted plane. Contact-zone decomposition is measured as mIoU against Split4D's reported 0.893 on Multi-Human. We also report PSNR and LPIPS, and frames per second on a fixed GPU.
A second decisive check asks whether the static/dynamic fix is real. We count floater Gaussians created within the contact band after lift-off, and compare unmodified Ex4DGS against our identity-based rule.
The project succeeds if continuous-time sliding drops substantially while PSNR and frame rate are unchanged. It fails informatively if the dense soft penalty does as well, which would show that structure is not needed.
References
- Fully Explicit Dynamic Gaussian Splatting
- SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video
- AD-GS: Object-Aware B-Spline Gaussian Splatting for Self-Supervised Autonomous Driving
- Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control
- Shoe Style-Invariant and Ground-Aware Learning for Dense Foot Contact Estimation
- Split4D: Decomposed 4D Scene Reconstruction Without Video Segmentation
- Contact-guided Real2Sim from Monocular Video with Planar Scene Primitives
- BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting
- FlashAvatar: High-Fidelity Head Avatar with Efficient Gaussian Embedding
- Improving 3D Foot Motion Reconstruction in Markerless Monocular Human Motion Capture
- Capturing and Inferring Dense Full-Body Human-Scene Contact
- LAGOON: Language-Guided Motion Control