Footprints as Rulers: Metric Scale from Gravity-Timed Stance Contacts in Monocular Video

A single moving camera can reconstruct a scene only up to an unknown global scale, and that scale drifts over long videos. The usual fix is to use a person in the frame as a ruler: people are roughly 1.7 m tall, so the scene can be sized to match. That ruler has a built-in error. A short person who is close and a tall person who is far away look the same, so any error in the assumed height passes one-for-one into the scale estimate. We propose a second, independent ruler that walking people already carry: gravity. Each time a foot is planted, it briefly becomes a fixed point in the scene. The time between footfalls is measured in real seconds, because the frame rate is known. Walking is a pendulum-like motion governed by gravity, so the relationship between stride timing and stride geometry fixes absolute size without any assumption about how tall the walker is.

The core move is to stop treating the static-scene assumption as something humans only violate. A foot in stance obeys it for a few hundred milliseconds. We add these stance footprints to the static reconstruction as landmarks. We read their spacing and timing to get a gravity-based scale estimate, and fuse that per trajectory segment with the usual body-size estimate. Interpenetration and floating serve two jobs: they decide when a foot really touches the reconstructed surface, and they provide a held-out check on whether the recovered scale is physically consistent.

Core Claim

Anchoring footfalls on the reconstructed static surface lets us measure stride length and leg length in the reconstruction's arbitrary units, and stride period in seconds. Consider three quantities: the scene-unit stride length $\ell_s$, the scene-unit leg length $L_s$ from the fitted body mesh, and the stride period $T$. The ratio $\ell_s/L_s$ does not depend on scale. Through a population gait relation $\ell/L = F(\mathrm{Fr})$, where $\mathrm{Fr} = v^2/(gL)$ is the dimensionless walking speed and $g$ is gravity, this ratio determines $\mathrm{Fr}$. Substituting $v = s\,\ell_s/T$ and $L = s\,L_s$ then gives the metric scale factor $s$ in closed form:

$$ s \;=\; \frac{F^{-1}(\ell_s/L_s)\; g\, T^2\, L_s}{\ell_s^{2}}. $$

This estimate never uses a height prior. Its error comes from how much individual gait varies around $F$, which is a different error source from anthropometric variance. We claim that fusing the two estimates in log-scale produces metric scale whose error is nearly flat as subject height departs from the population mean. Height-only rulers cannot do this. We further claim the fused estimate reduces long-range scale drift when it is applied per segment at each stance event.

The claim can fail. If walkers vary their speed so much that $F^{-1}$ is as uncertain as the height prior, the gravity ruler adds little. We test that directly; see Proposed Experiments.

Relation to Prior Work

Several systems already use the human body as a ruler, and they motivate this proposal because they inherit its weakness. HSfM scales unit-less pointmaps into meters using the SMPL-X distribution of body dimensions. ODHSR treats the body's invariant proportions as a scale landmark inside an online system with separate static-scene and human Gaussians. Metric from Human inpaints people into images to calibrate a depth head. All three put anthropometric error directly into scale, which is the failure that Lin and the age-aware models of Liu et al. point to. We keep their body ruler as one of our two estimates and add an estimate whose error does not depend on it.

WHAC is the closest work. Its MotionVelocimeter learns metric velocity from canonicalized motion and aligns scaleless visual odometry to it. It reports a scale factor near 1.1, where plain odometry drifts to 5.0 or more. WHAC reconstructs no scene, so its gait cue is inferred from pose alone and is never tied to where feet actually land. We make the gait cue explicit and physically grounded: stride length is measured between reconstructed footprints on the static map, not regressed. The same footprints also serve as static landmarks for the reconstruction.

Scene-contact methods constrain how contact is modeled. PROX and HSR penalize human–scene overlap with signed distance fields, which are functions giving each point's distance to a surface, negative inside it. PROX needs a pre-scanned scene, and HSR uses slow implicit optimization. PhySIC reports that aggressive penetration penalties push bodies into floating. Crowd4D shows that non-planar terrain needs an explicit interaction surface, and it takes about four hours per sequence. From ContactOpt we borrow the idea that a small tolerated penetration band gives stable gradients. We use it for a narrower purpose: deciding which frames count as stance on a noisy monocular surface, not solving full-body contact optimization.

Masking-based dynamic reconstruction warns against treating the human purely as clutter. RobustSplat and Split4D separate moving content from static background with no notion of scale or contact. Our stance footprints are the one part of a moving person that a static reconstruction should keep.

Research Program

We build on an ODHSR-style online pipeline: static-scene Gaussians plus an SMPL-driven human, initialized from monocular depth and 2D keypoints. We add three pieces.

The first is a stance detector. It flags a foot as planted when three conditions hold at once: the foot's world-frame velocity, in scene units normalized by leg length, falls below a threshold; the sole vertices lie within a tolerance band of the static surface, allowing a few centimeters of penetration for shoe and terrain compliance in the spirit of ContactOpt; and the foot does not float above the surface. Each stance event $k$ yields a footprint $p_k$, where the foot's ray meets the static map, together with a timestamp $t_k$. Because $p_k$ is a 3D point on the reconstructed surface, stairs and slopes need no ground-plane assumption.

The second is the pair of scale estimates. The gravity estimate applies the equation above to each same-foot stride. The gait relation $F$ and its spread are fitted on training portions of metric-ground-truth walking data (EMDB and SLOPER4D training splits) and checked on held-out subjects. The body estimate uses the contact-fixed depth of the person to measure body height in scene units and compares it with the SMPL-X population prior, as HSfM and ODHSR do.

The third is fusion. The two estimates are combined as inverse-variance-weighted Gaussians in $\log s$. The scale is modeled as piecewise-smooth over time, with a random-walk prior between stance events, so each footfall corrects accumulated drift the way a landmark does in SLAM. When several people are present, the scene scale is shared while heights stay individual. This lets outliers such as children be flagged instead of corrupting the scale. Floating and penetration of non-stance frames are held out entirely from estimation and used only for evaluation.

Baselines

The naive reference point is monocular visual odometry with no metric anchor, aligned to ground truth only by rotation and translation. This shows how large scale error and drift are when nothing is done. WHAC reports that such odometry drifts to scale factors of 5.0 or higher on EMDB2.

The strongest existing competitor is WHAC. It is the only cited method that already turns human motion into metric scale in moving-camera video, and it reports WA-MPJPE of 142.2 mm on EMDB2 (world-aligned mean per-joint position error), against 165.9 mm for WHAM. We run WHAC's released model on identical sequences with the same visual-odometry front end and the same keypoint and body estimators. It gets no scene-specific tuning; ours gets only a threshold search on the training split. Beating it means lower median log-scale error and lower WA-MPJPE on held-out EMDB sequences. It also means a clearly flatter dependence of scale error on subject height.

The strongest body-ruler competitor is ODHSR. It reports WA-MPJPE of 175 mm and camera ATE of 8.4 cm on EMDB; ATE is absolute trajectory error, the distance between estimated and true camera positions. It shares our backbone, so we implement our additions on top of it, which makes the comparison a clean add-on test.

A feed-forward metric depth reference uses Metric3D v2, which reports zero-shot AbsRel of 0.047 on NYUv2 but needs a focal length. We run it once with true intrinsics and once without, and compute scale by median ratio to the reconstruction. Beating it in the no-intrinsics setting is required. Matching it in the with-intrinsics setting is acceptable.

Two ablations isolate the key ingredient. The first uses only the body estimate with our stance-based contact, which is essentially an online Crowd4D/HSfM-style ruler. The second uses only the gravity estimate. A third ablation computes stride from pose-regressed foot positions instead of footprints on the static map. This tests whether anchoring on the scene matters or whether WHAC-style pose-only gait would do just as well.

Why not just…?

Why not just use WHAC's learned velocity, since it already encodes gait? This is the main threat. A network trained on enough motion may already capture the Froude relation implicitly. The difference is that WHAC regresses velocity from pose with no scene, so foot sliding and depth errors enter unchecked. We measure stride between footprints fixed on static geometry, where each footprint is itself a static landmark. The pose-regressed-stride ablation and the head-to-head comparison against WHAC settle this. If footprint-anchored gait does not beat WHAC's velocity, the novelty collapses to adding scene contact to WHAC, and we would report that.

Why not just rely on a metric depth foundation model? Metric3D v2 depends on focal-length metadata, and Metric from Human documents scene-dependent scale in feed-forward models. Neither model tracks drift over a long trajectory. The with- and without-intrinsics runs show whether the threat is real. We expect the foundation model to be competitive when intrinsics are known and to degrade without them.

Why not just use body height plus contact, as Crowd4D and HSfM do? Contact removes depth ambiguity but not height ambiguity: the person's height in scene units is measured exactly, yet it is converted to meters through a prior with population spread. The height-stratified experiment below tests whether the gravity ruler removes this confound.

Why not just fit a ground plane? Crowd4D shows that plane assumptions make walkers sink into stairs or float. Our footprints are ray–surface intersections, and the VirtualCrowd terrain split tests this.

Proposed Experiments

The decisive experiment is height-stratified. We render synthetic walkers whose SMPL-X shapes span child to very tall adult into scenes with known metric geometry, and add EMDB subjects binned by height. We plot log-scale error against the subject's deviation from mean height. Body-ruler baselines (ODHSR, the body-only ablation) should show error that grows roughly linearly with deviation. Our method should stay flat. We also report the gravity-only spread, which directly measures how noisy $F$ is.

Drift is tested on SLOPER4D's kilometer-scale sequences and on EMDB walks. We report scale error per 100 m of travel and ATE after rotation-and-translation alignment only, without a scale correction, so scale errors cannot be hidden. SLOPER4D reports LiDAR fusion with 13% lower trajectory error than image-only methods, but no absolute value is given. For absolute metric scale error on EMDB or SLOPER4D, no reported result was found on Lacuna, so we compute it for all baselines ourselves.

Terrain and crowds are tested on VirtualCrowd. Crowd4D reports PPDS 91.43 there; PPDS is a population-density score for how well the crowd is placed in the scene. We report PPDS, MPJPE, and runtime against Crowd4D's roughly four hours per sequence.

Physical consistency is a held-out check. Foot penetration depth and floating distance on non-stance frames, plus foot skate, are never used during estimation. If scale is wrong, these should rise systematically, so they test whether the recovered metric scale is physically right.

The proposal fails in two cases. The first is if gait spread on held-out subjects exceeds anthropometric spread by enough that fusion does not flatten the height-stratified curve. The second is if pose-regressed stride matches footprint-anchored stride.