Plücker Ray-Distance Factor in SLAM

• Plücker Ray-Distance Factor is a geometric constraint that uses Plücker coordinates to measure the perpendicular distance from 3D landmarks to world-anchored metric rays.
• It resolves scale ambiguity in monocular SLAM by penalizing inconsistent landmark scaling, thereby converting scale from a free parameter into an observable quantity.
• The factor incorporates uncertainty weighting through DSUG and demonstrates strong empirical performance by reducing metric drift, as evidenced in PRISM-SLAM benchmarks.

Searching arXiv for the specified paper and closely related Plücker-based line/ray distance work. arXiv search query: "PRISM-SLAM Plücker Ray-Distance Factor PRISM-SLAM (Im, 19 May 2026)" The Plücker Ray-Distance Factor is a Plücker-coordinate-based geometric constraint that, in its explicit modern usage, denotes the point-to-ray residual introduced in PRISM-SLAM for scale-aware monocular metric SLAM. It represents vision-foundation-model (VFM) predictions as metric 3D rays anchored in the world, and penalizes the orthogonal distance from reconstructed landmarks to those rays. In PRISM-SLAM, this factor is the mechanism that anchors monocular observations in an absolute metric frame, resolves scale ambiguity and scale drift, and makes global metric scale locally Fisher-identifiable. More broadly, related research uses Plücker coordinates to encode line validity, Grassmannian distance, or bilinear ray relations; these constructions clarify that “Plücker Ray-Distance Factor” can refer either to a specific SLAM residual or, more generally, to a family of scale-aware or incidence-aware constraints defined on Plücker line representations (Im, 19 May 2026).

1. Definition and conceptual role

In monocular SLAM, pure pinhole-camera geometry is defined only up to a global scale. PRISM-SLAM recalls the standard reprojection cost

$$E = \sum_{i,j} \rho \left( \left\| p_{ij} - \pi\big(K (R_i X_j + t_i)\big) \right\|^2_{\Sigma_{ij}} \right),$$

together with the invariance

$$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$

so the Hessian has a null direction in scale. Scale is therefore unobservable and can drift. The Plücker Ray-Distance Factor is introduced to break exactly this invariance by coupling map landmarks to VFM-derived metric rays fixed in a globally consistent coordinate system (Im, 19 May 2026).

The factor does not inject VFM output as a simple scalar depth residual. Instead, it represents VFM information as metric 3D rays in Plücker coordinates, and measures the orthogonal distance from SLAM map points to those rays. Because the rays are metric and world-anchored, a global rescaling of the SLAM structure moves landmarks away from their associated rays and strictly increases the residual. This supplies explicit scale gradients, eliminates the rank-deficient null-space of reprojection-only monocular SLAM, and converts scale from a gauge freedom into an observable quantity.

This role is central to the architecture of PRISM-SLAM. The factor is described as the component that anchors monocular observations in an absolute, globally consistent metric coordinate system, resolves scale ambiguity and scale drift, and makes the metric scale Fisher-identifiable. A plausible implication is that the term names not merely a residual formula, but a specific observability mechanism: a Plücker-encoded metric prior that injects curvature into the scale direction of the likelihood.

2. Mathematical formulation

For each camera $i$ with pose $T_i=(R_i,t_i)$, PRISM-SLAM constructs a world-frame metric ray from the VFM prediction. Let $d_i\in\mathbb{R}^3$ be the ray direction and let $t_i\in\mathbb{R}^3$ be the camera center in the world frame. The ray is represented by Plücker coordinates

$L=(d_i,m_i), \qquad m_i=t_i\times d_i.$

Here $d_i$ is the line direction and $m_i$ is the Plücker moment, encoding the line’s offset from the origin. The backend jointly optimizes camera poses $T_i\in SE(3)$, landmarks $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$0, and a global scale state estimated in a separate log-domain Kalman filter, while the graph itself remains metric (Im, 19 May 2026).

For a ray from camera $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$1 and a landmark $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$2, the Plücker Ray-Distance residual is

$$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$3

The numerator is the magnitude of the moment of $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$4 with respect to the line, geometrically equal to the perpendicular distance from the point to the infinite line. The residual therefore penalizes the Euclidean distance from the reconstructed point to the VFM-anchored ray.

The factor enters the posterior probabilistically through a likelihood term

$$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$5

with information form

$$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$6

where $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$7 is the Dynamic Scene Uncertainty Gating (DSUG) weight at pixel $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$8. The overall posterior factorizes over standard reprojection factors, ray-distance factors, and additional depth or pose priors and loop-closure constraints, and is optimized by nonlinear least squares using Levenberg–Marquardt.

This formulation is “ray-grounded” in a strict geometric sense. The residual is a 3D point-to-line distance rather than a 2D image-plane error, and the Plücker representation encodes both direction and position in a form that is well behaved under rigid motion while remaining sensitive to scale changes.

3. Scale observability and metric gauge fixing

The defining analytic property of the factor is its dependence on global scale. If a global scale $$\pi\big(K (R_i (s X_j) + s t_i)\big) = \pi\big(K (R_i X_j + t_i)\big), \quad \forall s>0,$$9 is applied to the landmarks, $i$0, while the VFM-derived ray remains fixed in metric space, then the residual becomes

$i$1

For generic configurations this quantity changes with $i$2, and its derivative with respect to $i$3 is generally non-zero. In Fisher-information terms, the log-likelihood acquires non-zero curvature in the scale direction, so scale becomes locally observable and Fisher-identifiable (Im, 19 May 2026).

This distinguishes the Plücker Ray-Distance Factor from both classical reprojection and simple scalar depth residuals. The reprojection model is invariant to scale by construction. Scalar depth priors constrain depth along individual pixel rays, but the PRISM-SLAM description emphasizes that they do not exploit the full 3D line geometry or produce the same clean scale-sensitive information structure. By contrast, the Plücker factor ties landmarks to world-anchored metric rays and forces any inconsistent rescaling to manifest as a physical deviation from the ray.

The experimental consequences are reported in metric $i$4 evaluation. On TUM fr1/xyz, PRISM-SLAM reports Sim(3) / $i$5 ATE of $i$6 cm, corresponding to only $i$7 scale error. On TUM fr3 static, the reported values are $i$8 cm for sit and $i$9 cm for walk. The small gap between oracle-aligned Sim(3) and strict $T_i=(R_i,t_i)$0 alignment is presented as evidence that the system no longer behaves as a free $T_i=(R_i,t_i)$1 gauge system, but effectively as an $T_i=(R_i,t_i)$2 system with fixed metric scale.

Ablation further isolates the role of the factor. Removing the Plücker Ray Factor increases mean ATE from $T_i=(R_i,t_i)$3 cm to $T_i=(R_i,t_i)$4 cm, a degradation of $T_i=(R_i,t_i)$5 cm, which is larger than the increases reported for removing DSUG, log-domain Kalman scale filtering, or WLS. Within the reported study, this identifies the factor as the dominant contributor to accurate metric scale.

4. Integration with uncertainty and dynamic-scene robustness

PRISM-SLAM couples the factor to VFM uncertainty through Dynamic Scene Uncertainty Gating. For each pixel $T_i=(R_i,t_i)$6, the system constructs a hybrid uncertainty

$T_i=(R_i,t_i)$7

and then a soft gate

$T_i=(R_i,t_i)$8

Here $T_i=(R_i,t_i)$9 is derived from DA3’s spatial confidence and $d_i\in\mathbb{R}^3$0 is a pose-compensated temporal depth-variance term. The resulting weight modulates the information of the ray-distance factor and any associated depth factors derived from the same pixel (Im, 19 May 2026).

This construction is explicitly motivated by dynamic-scene failure modes. If moving objects or erroneous VFM predictions were injected directly into the backend, the resulting ray factors could pull the map toward incorrect geometry and distort scale. DSUG addresses this by assigning strong weights to static, confident regions and down-weighting dynamic or uncertain regions. In static regions, $d_i\in\mathbb{R}^3$1 is small and $d_i\in\mathbb{R}^3$2, so the metric anchoring remains strong. In dynamic or uncertain regions, $d_i\in\mathbb{R}^3$3 is large and $d_i\in\mathbb{R}^3$4, so the corresponding constraints are softened rather than removed through hard semantic masking.

The empirical dynamic-scene results are reported on BONN Dynamic. The full system achieves $d_i\in\mathbb{R}^3$5 cm on balloon, $d_i\in\mathbb{R}^3$6 cm on balloon2, $d_i\in\mathbb{R}^3$7 cm on pers_trk, and $d_i\in\mathbb{R}^3$8 cm on balloon_trk. The same ablation study reports that removing DSUG raises mean $d_i\in\mathbb{R}^3$9 ATE from $t_i\in\mathbb{R}^3$0 cm to $t_i\in\mathbb{R}^3$1 cm, a $t_i\in\mathbb{R}^3$2 cm increase. This establishes that the effectiveness of the Plücker Ray-Distance Factor in dynamic scenes depends not only on the residual definition, but also on uncertainty-aware weighting inside the factor graph.

5. Related Plücker-based distance notions in adjacent literatures

The term Plücker Ray-Distance Factor is specific in PRISM-SLAM, but several adjacent literatures define related quantities on Plücker line representations. These are not interchangeable, and the distinctions matter.

In Plücker correction, a noisy 6D vector $t_i\in\mathbb{R}^3$3 is projected onto the Klein quadric $t_i\in\mathbb{R}^3$4, yielding the closest valid Plücker line under the Euclidean norm in $t_i\in\mathbb{R}^3$5. The objective is

$t_i\in\mathbb{R}^3$6

which is equivalent to minimizing

$t_i\in\mathbb{R}^3$7

This defines an algebraic distance from an arbitrary 6D vector to the manifold of valid Plücker lines, and the paper provides a simple closed-form global solution that does not require singular value decomposition (Cardoso et al., 2016). This is a Plücker-space correction distance, not PRISM-SLAM’s world-space point-to-ray distance.

In RayPE, the key Plücker quantity is the reciprocal product

$t_i\in\mathbb{R}^3$8

for rays $t_i\in\mathbb{R}^3$9 and $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$0. This bilinear form vanishes iff the two lines are coplanar, is $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$1-invariant, and scales linearly with global translation scale through the moments. The paper is explicit that this reciprocal product is not a metric in the sense of Euclidean distance between lines; it is more accurately a ray-incidence or coplanarity factor, although in attention it behaves as a “ray-geometry factor” or “distance-like factor” (Yin et al., 25 Jun 2026). This corrects a common misconception: not every scalar built from Plücker coordinates is a distance.

In distance optimization on the Grassmannian of lines, a line is represented by a rank-2 skew-symmetric Plücker matrix $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$2, which is mapped to a projection matrix by squaring and normalization,

$L=(d_i,m_i), \qquad m_i=t_i\times d_i.$3

A natural scale-invariant alignment factor between Plücker rays $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$4 and $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$5 is

$L=(d_i,m_i), \qquad m_i=t_i\times d_i.$6

from which chordal distance follows as

$L=(d_i,m_i), \qquad m_i=t_i\times d_i.$7

The paper does not introduce the term “Plücker Ray-Distance Factor,” but it explicitly develops a scale-invariant distance structure on Plücker rays through projection coordinates and Grassmann metrics (Friedman et al., 30 Jan 2026).

In lifted Gabidulin list decoding, Plücker coordinates are used to separate code constraints from distance constraints. The paper does not use the phrase “Plücker Ray-Distance Factor,” but its ball equations in Plücker coordinates define a distance-only component in projective Plücker space, independent of code structure. This suggests an abstract interpretation of a Plücker distance factor as the subset of Plücker rays satisfying ball constraints around a received subspace (Trautmann et al., 2013).

Taken together, these literatures indicate that the phrase can denote at least three distinct objects: a point-to-ray residual in Euclidean 3D space, a correction distance to the Klein quadric in $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$8, or a scale-invariant Grassmannian distance or incidence factor on projective Plücker rays.

6. Implementation profile, empirical role, and limitations

In PRISM-SLAM, the factor is implemented inside backend Metric Graph Optimization in a local window of $L=(d_i,m_i), \qquad m_i=t_i\times d_i.$9 keyframes with 15 Levenberg–Marquardt iterations, followed by a global bundle adjustment after loop closure. Huber robust kernels are applied with $d_i$0 and $d_i$1. The residual is differentiated with respect to both $d_i$2 and $d_i$3, and the paper notes that the Jacobians are straightforward via the chain rule. The factor uses a compact 3D residual per point rather than a heavy volumetric representation, and the full RGB-only system is reported to run at $d_i$4 FPS (Im, 19 May 2026).

The broader significance claimed for the factor is methodological. It provides a pattern for integrating VFM priors into SLAM by expressing them as probabilistic 3D line constraints with uncertainty, rather than as deterministic depth targets. The paper further states that the approach is compatible with other VFMs, including Metric3D, UniDepth, and VGGT, provided they can supply metric depths or rays. This suggests that the factor is a reusable design motif for structured Bayesian SLAM systems that aim to operate directly in metric $d_i$5 without post-hoc Sim(3) alignment.

The reported limitations are equally specific. The factor relies heavily on DA3’s metric correctness and cross-view consistency; if the VFM’s metric scale is biased or inconsistent across scenes, the injected Plücker rays become biased metric priors. In highly dynamic scenes, some erroneous rays may survive DSUG and temporarily distort metric scale. The construction also assumes sufficiently accurate camera poses for transforming rays into world coordinates; severe early-stage pose errors can degrade the quality of the ray anchors. Finally, the paper does not provide a full analytic observability analysis in terms of explicit Fisher information matrices, relying instead on geometric argument and empirical evidence.

Under the usage established by PRISM-SLAM, then, the Plücker Ray-Distance Factor is best understood as a probabilistic, scale-sensitive, point-to-ray constraint defined by

$d_i$6

with uncertainty-weighted information and direct implications for observability. In the broader Plücker literature, related constructions show that the same terminology can be extended—carefully and with domain-specific distinctions—to Euclidean correction on the Klein quadric, Grassmannian distance optimization, and bilinear ray-incidence modeling.