Anchored Modal Curves: Joint Camera and Shape Recovery for Moving Thin Structures

Swaying power lines, branches, fences shaking in the wind, and dangling cables are hard for 3D reconstruction in two separate ways. They are thin: each covers only a few pixels and has almost no texture, so point-feature trackers and learned depth networks miss them. They also move, which breaks the assumption behind camera tracking that the world stays still while the camera moves. Existing systems handle one problem or the other. Vid2Curve recovers camera motion and wire geometry together from a handheld video, but it assumes the wires never move. 4D3R recovers camera motion in dynamic scenes, but it depends on a textured static background and on learned depth and segmentation models that break down on thin structures.

This note proposes to exploit a physical fact the literature has not used: most real thin structures are attached at their ends and bend in only a few smooth ways. Each structure is modeled as a 3D curve with fixed attachment points plus a small set of bending patterns, and the camera trajectory is solved jointly with the per-frame weights on those patterns. Camera motion then becomes a global 6-degree-of-freedom change shared by every curve, while deformation is local, small in dimension, and zero at the anchors. That separation is what makes "did the camera move or did the wire move?" answerable from the curves alone, with no textured background required.

Core Claim

Consider a monocular RGB video of a scene dominated by anchored thin structures, with little usable static texture. The claim is that jointly optimizing camera poses and an anchored, low-dimensional curve deformation model recovers both trajectory and time-varying geometry. Two alternatives should fail in this setting: static curve-based joint optimization, which treats motion as outliers, and dynamic Gaussian methods that estimate pose from the background.

The key ingredient is a structural restriction on the deformation. Curve $k$ at time $t$, with arc-length parameter $s\in[0,1]$, is written as

$$ x_k(s,t) = \bar{x}_k(s) + \sum_{m=1}^{M} a_{k,m}(t)\,\phi_m(s),\qquad \phi_m(0)=\phi_m(1)=0, $$

Here $\bar{x}_k$ is the rest shape of the curve. The $\phi_m$ are $M$ smooth bending patterns that vanish at the anchor points, like the vibration modes of a cable fixed at both ends. The $a_{k,m}(t)$ are per-frame weights. For a structure with one free end, such as a branch, the patterns vanish only at the root.

Camera poses $(R_t, T_t)$ and all curve parameters are fit by minimizing a distance between each projected curve and a 2D curve map detected in each frame. That distance is measured only perpendicular to the detected curve, since motion along a featureless wire cannot be observed. A small penalty on the weights $a$ and a temporal smoothness term complete the objective. No photometric term enters the pose or geometry stage. Appearance is added afterwards by attaching cylindrical Gaussians to the recovered curves, following DGH.

The expectation that a small $M$ suffices is a hypothesis to be tested, not a given. It rests on the observation that slender structures driven by wind or contact store most of their motion energy in low-frequency bending.

Why This Is Different From What Exists

Three strands of prior work each cover part of the problem.

Vid2Curve motivates and supplies the method. It shows that 2D skeleton curves can replace point features for joint camera and wire estimation, and that classical multi-view stereo fragments wire-like objects. Its dynamic-programming curve matching and camera registration, however, assume a rigid scene. We keep its curve-distance registration and replace the rigid curve network with the anchored modal model.

The dynamic, pose-free Gaussian systems motivate the gap. 4D3R uses masked PnP on static anchors found by MonST3R and SAM-2. DynSUP assumes piecewise-rigid object motion. Mono4DGS-HDR uses an as-rigid-as-possible constraint. Free-DyGS and Mono-STAR need depth input. All five rely on bulky surfaces or dense priors that the research file identifies as unreliable on sub-pixel geometry.

Curve and strand primitives supply the rendering layer. Curve-Aware Gaussian Splatting, HairGS, and DGH show that curve-attached Gaussians represent thin geometry well, but all require known camera poses, and the first two are static.

KeyGS constrains the design. It shows that under joint optimization, photometric Gaussian fitting absorbs pose error into floaters, which then corrupt the pose gradients. This is why our pose stage has no free-floating primitives at all: every residual must be explained either by camera motion or by a bounded bending of an anchored curve.

RobustSplat is another constraint. Its delayed-densification transient filtering discards moving content, and at low resolution it would discard thin wires with it.

Research Program

The system has three parts.

The first is a 2D front end. It produces per-frame curve maps with orientation, which are needed for the perpendicular-distance loss. We start from an off-the-shelf edge or ridge detector and measure how sensitive the whole pipeline is to its errors, because Vid2Curve's dependence on clean masks is a known weakness.

The second is the joint optimizer. Initialization comes from a short, near-rigid window, where the weights $a$ are set to zero and the method reduces to Vid2Curve-style registration. The window then slides forward, adding frames and solving jointly for new camera poses and per-curve weights. Anchor positions are estimated as part of the rest shape and are shared across all time. Bending patterns are first a fixed analytic basis (sine modes on arc length for two-anchor curves, cantilever modes for one-anchor curves). A second variant learns a per-curve basis by factorizing an initial free-deformation fit. Deciding whether the analytic basis is enough is part of the study.

The third is an identifiability analysis. Anchors are static, so a family of anchor points seen from enough viewpoints pins down the camera in the same way static points do in structure-from-motion. The analysis will make precise when the camera stays recoverable as the fraction of anchors in view drops. We will also test the method empirically on the failure case it does not claim to solve: free-floating thin structures with no visible anchor.

Baselines

The naive reference is Vid2Curve run unchanged on dynamic sequences. We also run a stronger naive variant that adds robust outlier rejection, down-weighting curve segments with persistently high reprojection residuals. This is the cheapest way to "make static methods tolerate motion." We beat it if camera error stays low and geometry is recovered for the moving segments, which this variant discards by construction. Vid2Curve reports reconstruction error "consistently less than 0.001" and projection error "under 0.003 pixels" on its static synthetic tests. Our method should match these on static scenes, which is a check that nothing regresses.

The strongest existing joint dynamic method is 4D3R. Its reported result is "up to a 1.8dB PSNR improvement" over SC-GS and RoDynRF on dynamic monocular video. It gives no thin-structure numbers: no reported result found on Lacuna for dynamic thin-structure pose or geometry accuracy. We run the released pipeline on identical frames with its default MonST3R and SAM-2 front end. Our method gets the same GPU budget and a comparable per-scene optimization time, measured against 4D3R's reported roughly 50-minute HyperNeRF training. We beat it if camera error is lower and curve-geometry accuracy higher on thin-dominated scenes, while novel-view quality is not worse on standard HyperNeRF scenes.

The decoupled baseline estimates poses from the background with 4D3R's pose stage or MonST3R. It then fits our identical anchored modal curves with those poses frozen. This isolates whether joint optimization matters.

The ablations isolate the key ingredient. The first replaces the anchored modal deformation with free per-point displacement plus a smoothness penalty. The second keeps the patterns but removes the zero-at-anchor constraint. The third adds a photometric Gaussian term into the pose stage. An oracle run with ground-truth poses bounds how much of the geometry error comes from pose.

Why not just…?

Why not just estimate poses from the static background, then fit dynamic curves with known poses? This is the most serious threat. When a textured static background fills much of the frame, it will likely work about as well as our method. The decoupled baseline above tests it directly, and we will report results binned by the fraction of static textured pixels in view. The claim only holds if our method wins clearly in the low-texture bins, such as sky behind power lines or a wire sculpture on a plain backdrop. There, Vid2Curve already reports that point-feature pipelines fragment, and 4D3R's masked PnP has nothing to lock onto. If the decoupled pipeline matches us across all bins, joint optimization is unnecessary and we will say so.

Why not just treat moving segments as outliers in Vid2Curve? Outlier rejection recovers the camera only while the static portion dominates. It also returns no geometry for exactly the parts that move. We expect camera error to rise sharply as the moving fraction grows. The robust Vid2Curve baseline measures that curve.

Why not just use a general dynamic Gaussian model with free deformation and joint pose refinement? Unconstrained deformation can absorb global camera motion. KeyGS shows that free primitives turn pose error into floaters. The free-displacement ablation tests whether the low-dimensional anchored restriction, not the curve representation alone, is responsible for the gain.

Proposed Experiments

No existing benchmark in the file has moving thin structures together with ground-truth poses. The primary benchmark is therefore a new synthetic set, built in the spirit of the XPBD-simulated data DGH trained on. It will contain simulated cables, branches, fences, and rigging under wind and contact, rendered along handheld-style trajectories. Ground truth includes camera poses and per-frame curves, and each scene varies how much textured background is visible.

The metrics are absolute trajectory error, curve precision, recall, and F-score at fixed distance thresholds as used for strands in HairGS, and per-frame geometric error of the moving segments.

The real-data evaluation has two parts. The first is a small set of handheld captures of swaying outdoor structures, with poses from a motion-capture system following the MultiCamRobolab protocol of Dense Dynamic Scene Reconstruction and Camera Pose Estimation from Multi-View Videos. That paper reports an absolute translation error of 0.005 m on MultiCamVideo, which sets the precision scale we should aim for. The second is HyperNeRF and MPI Sintel, used only to confirm that the curve-attached rendering layer does not break general dynamic scenes. There we compare PSNR, SSIM, and LPIPS against 4D3R, restricted to regions containing thin structures.

For context, DGH reports PSNR 27.01 with calibrated cameras on its own hair data. This is a known-pose reference in a different domain, not a directly comparable target.

Three results would decide the question. The first is camera error and moving-segment accuracy plotted against visible static texture: the method is worthwhile only if it clearly separates from the decoupled baseline at low texture. The second is the free-deformation ablation, which shows whether the anchored low-dimensional model is what breaks the camera-versus-motion ambiguity. The third is the fit quality as a function of the number of bending patterns $M$. If real structures need so many patterns that the model approaches free deformation, the core hypothesis is wrong.