Scene Correspondence Priors

• Scene correspondence priors are structured biases that guide matching by favoring semantically, geometrically, or functionally plausible associations across images, point clouds, and time steps.
• They integrate engineered and learned regularities—such as disparity offsets, semantic graphs, and latent diffusion—to constrain correspondence estimation before ambiguous scenarios arise.
• Recent research shows these priors boost performance in applications like 3D pose recovery, change detection, navigation, and scene augmentation, especially under challenging conditions.

Searching arXiv for papers related to scene correspondence priors and the cited works. I’ll look up the named papers and related terms on arXiv to ground the article in current literature. Scene correspondence priors are prior structures that bias correspondence estimation toward semantically, structurally, or geometrically plausible associations across images, point clouds, time steps, scene entities, or object instances. In current arXiv literature, the term does not denote a single formalism. It spans early cross-view token interaction for RGB-D object pose estimation, preferred disparity-surface slants in stereo, overlap and occlusion priors in unaligned change detection, semantic and functional priors for navigation and scene parsing, and learned priors over dense pixel-to-3D assignments or motion fields (Chen et al., 20 Jun 2026, Scharstein et al., 2017, Liu et al., 14 Sep 2025, Yang et al., 2018, Bian et al., 14 Oct 2025). The common thread is that correspondence is treated not as an unconstrained matching problem, but as one constrained by learned or engineered regularities of scene structure.

1. Conceptual scope and taxonomy

Across the literature, a correspondence prior specifies what kinds of matches should be preferred before, during, or after geometric decoding. In some formulations, the prior is explicitly geometric: a stereo method prefers disparity transitions consistent with a slanted surface rather than a fronto-parallel patch, or a change detector restricts comparison to reprojected, overlapping, and visible regions. In others, the prior is semantic or relational: a navigation agent infers that mugs correspond functionally to coffee machines and cabinets, or a scene parser infers that labels in one spatial region constrain labels in another. A third family treats priors as properties of a learned representation: a scene-coordinate regressor is regularized by a depth or diffusion prior over predicted 3D points, and a runtime-optimized MLP acts as an implicit prior over a continuous scene-flow field (Scharstein et al., 2017, Yang et al., 2018, Zhang et al., 2018, Bian et al., 14 Oct 2025, Li et al., 2021).

Domain	Correspondence target	Prior form
Single-reference RGB-D pose	3D-to-3D point matches	Cross-view semantic-structural-geometric prior
Stereo	Disparity transitions	Surface orientation prior
Semantic navigation	Target-context associations	Knowledge-graph semantic prior
Scene parsing / augmentation	Region-label or object-room relations	Spatial relational prior
SCR / relocalization	Pixel-to-3D assignments	Depth or diffusion prior
Dynamic point clouds / 4D scenes	Temporal point trajectories	Implicit flow prior or object view-synthesis prior

This diversity makes a narrow definition misleading. A common misconception is that scene correspondence priors must be geometric in the classical CV sense. The literature shows otherwise: some priors act on epipolar or reprojection structure, some on semantic compatibility, and some on the latent geometry induced by a model’s predictions (Yang et al., 2018, Bratulić et al., 12 Dec 2025).

2. Geometric priors over disparity, overlap, and visibility

A canonical geometric formulation appears in "Semi-Global Stereo Matching with Surface Orientation Priors" (Scharstein et al., 2017). Standard SGM uses a first-order smoothness term with zero cost at equal disparity and small cost for unit disparity difference, which encodes a fronto-parallel bias. SGM-P replaces this generic bias with a prior-specific bias by rasterizing a prior disparity surface and shifting the pairwise penalty so that the preferred local transition follows the prior slant. For a 2D prior surface $S$, the rasterized jump is $j_p = \hat{S}(p') - \hat{S}(p)$, and the modified smoothness becomes

$V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$

The prior therefore does not clamp disparity values; it redefines what local smoothness means. The 3D extension makes the offset disparity-dependent, which is necessary when multiple overlapping surface hypotheses or depth-dependent normal-induced slants exist. This paper is explicit that such priors are scene correspondence priors because they encode preferred local evolution of correspondence surfaces in disparity space rather than only generic regularity (Scharstein et al., 2017).

A second geometric formulation appears in unaligned scene change detection. "Leveraging Geometric Priors for Unaligned Scene Change Detection" introduces three priors derived from a GFM: a geometric correspondence prior, a visual overlap prior, and an occlusion prior (Liu et al., 14 Sep 2025). The method estimates $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$, reprojects a query pixel $p_1$ into the other view,

$$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$

restricts comparison to the overlap set

$$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$

and declares occlusion by depth inconsistency,

$$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$

This converts correspondence from unconstrained 2D appearance matching into a physically grounded sequence of tests: does the point reproject, does it lie in overlap, and is it visible. In the reported ablation, replacing optical flow with geometric matching raised PASLCD F1 from $0.094$ to $0.237$, and adding occlusion reasoning raised it further to $j_p = \hat{S}(p') - \hat{S}(p)$0, indicating that the prior’s value lies not only in better matching but in explicit reasoning about non-overlap and occlusion (Liu et al., 14 Sep 2025).

These works establish a strong geometric reading of scene correspondence priors: they are priors over admissible correspondence geometry, visibility, or surface evolution, injected directly into inference rather than learned only through end-task supervision.

3. Semantic, functional, and relational priors

A distinct lineage treats correspondence priors as semantic relations among scene entities rather than geometric matches. "Visual Semantic Navigation using Scene Priors" encodes object categories as graph nodes and Visual Genome relations as graph edges, then injects the resulting GCN representation into an A3C policy (Yang et al., 2018). In that formulation, the relevant correspondences are target object $j_p = \hat{S}(p') - \hat{S}(p)$1 context objects, unseen target $j_p = \hat{S}(p') - \hat{S}(p)$2 semantically related known objects, and visible scene context $j_p = \hat{S}(p') - \hat{S}(p)$3 likely hidden target locations. The graph uses $j_p = \hat{S}(p') - \hat{S}(p)$4 visible AI2-THOR categories, edges are added when a relation occurs more than three times in Visual Genome, and node features combine fastText embeddings with image classification scores from the current observation. The agent does not estimate metric geometric matches; it reasons over semantic/functional correspondence such as mug $j_p = \hat{S}(p') - \hat{S}(p)$5 cabinet near coffee machine or mango $j_p = \hat{S}(p') - \hat{S}(p)$6 fruit-like objects $j_p = \hat{S}(p') - \hat{S}(p)$7 fridge (Yang et al., 2018).

"Spatially Constrained Location Prior for Scene Parsing" provides an older but still precise relational formulation (Zhang et al., 2018). SCLP divides the image into $j_p = \hat{S}(p') - \hat{S}(p)$8 spatial blocks and estimates block-conditioned class co-occurrence statistics $j_p = \hat{S}(p') - \hat{S}(p)$9, thereby combining absolute and relative location priors. Local adjacency priors $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$0 complement this nonlocal block model. The prior acts by converting provisional superpixel labels into contextual votes: $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$1 Here the “correspondence” is probabilistic compatibility across spatially related regions. The same broad idea reappears in "SceneGen: Generative Contextual Scene Augmentation using Scene Graph Priors," which represents object-to-room, object-to-group, and orientation relations in a spatial scene graph and fits KDE models over relational feature vectors $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$2 and $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$3 per category (Keshavarzi et al., 2020). The output is a positional heatmap and, for asymmetric objects, an orientation likelihood for plausible augmentation in an existing scene (Keshavarzi et al., 2020).

A temporally weaker but related prior appears in "Using Image Priors to Improve Scene Understanding" (Schroeder et al., 2019). There the prior is simply an earlier frame $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$4 from the same driving scene, fused with the current frame $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$5 either at input, at the bottleneck, or at multiple decoder levels. The method does not use explicit geometric alignment, optical flow, or localization at inference time; the prior is a weak correspondence signal arising from temporal continuity and revisitation. The Decoder Prior improved IoU from $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$6 to $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$7, with the abstract reporting dynamic-class accuracy improving from $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$8 to $V_S(d_p,d_{p'}) = V(d_p + j_p, d_{p'}).$9 and static-class accuracy from $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$0 to $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$1 (Schroeder et al., 2019).

Taken together, these methods show that scene correspondence priors can be semantic and relational without being metric. They regularize where to search, what labels are compatible, and how contextual evidence should transfer across regions or time, even when explicit geometric correspondence is absent.

4. Cross-view priors for rigid objects and category-level semantics

In single-reference unseen object pose estimation, the correspondence bottleneck is explicit: the system must infer 3D-to-3D matches between a query RGB-D observation $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$2 and a single reference $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$3 of the same unseen rigid object. "Learning Cross-View Semantic Priors for Single-Reference Unseen Object Pose Estimation" argues that recent VFM-based pipelines largely treat image features as intra-view descriptors and defer cross-view reasoning until sparse geometric decoding, which is limiting under large viewpoint changes, sparse overlap, segmentation noise, occlusion, repeated geometry, and clutter (Chen et al., 20 Jun 2026). Its remedy is an early cross-view semantic prior built by $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$4 CVSI blocks over dense VFM tokens $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$5. The bidirectional cross-attention update for the query branch is

$$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$6

followed by a residual and FFN refinement. Two training-time constraints stabilize this prior: IVSP preserves the original intra-view token affinity structure, and RAGC enforces consistency of decoded point features in a reference-anchored coordinate frame. The final pose is recovered from learned correspondences by weighted SVD. On LM-O, TUD-L, and YCB-V under the predefined protocol, the method reports AR$$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$7 $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$8, mean $$\mathcal{K}, \mathcal{T}, \mathcal{D} = \mathscr{G}(I_1, I_2)$$9, compared with $p_1$0 for UNOPose and $p_1$1 for COG; with a DINOv3 backbone, adding only CVSI improves AR$p_1$2 from $p_1$3 to $p_1$4, and the full model reaches $p_1$5, while a shared-weight CVSI variant reaches $p_1$6. The gains are largest at severe viewpoint gaps, with $p_1$7 points improvement at both $p_1$8 and $p_1$9 relative rotation on YCB-V (Chen et al., 20 Jun 2026).

A related but category-level formulation appears in "Category-Level 3D Correspondence in Camera Space via Morphable Object Priors" (Sommer et al., 27 May 2026). Here the prior is a shared morphable object model per category rather than explicit pairwise correspondence supervision. Morpheus learns a category template as an SDF, an image-conditioned affine deformation field

$$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$0

and uses a pretrained 6D pose diffusion model to disentangle pose from deformation. Correspondence emerges because all instances are deformations of one canonical template with fixed topology and persistent vertex identity; a query point is projected to the query mesh, encoded as barycentric coordinates on a face, and transferred to the target mesh with the same face identity and barycentric coordinates. HouseCorr3D provides $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$1 test images, $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$2 train images, $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$3 categories, $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$4 instances, amodal labels, and symmetry annotations. On the full benchmark, Morpheus reports mean combined 3D [email protected] of $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$5, versus $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$6 for GenPose++ and $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$7 for MagicPony+GP++; on the six-category main table it reaches $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$8 in 3D modal and $$x_2^1 = R_{2\leftarrow 1}\, D_1(p_1)\, K_1^{-1}\tilde p_1 + t_{2\leftarrow 1}, \qquad p_2^1 = \Phi_{2\leftarrow 1}(p_1) = \pi(K_2 x_2^1),$$9 in 3D amodal (Sommer et al., 27 May 2026).

These object-centric works are narrower than full-scene priors, but they crystallize a core principle: correspondence becomes easier when semantic conditioning or canonical grounding is introduced before hard geometric disambiguation. The first paper does so by early cross-view token interaction; the second by persistent identity on a morphable category template.

5. Priors in implicit 3D, motion, and generative scene representations

A large recent literature treats correspondence priors as priors over the latent geometry induced by model outputs. "Scene Coordinate Reconstruction Priors" explicitly reframes scene coordinate regression as MAP estimation over dense pixel-to-3D correspondences (Bian et al., 14 Oct 2025). With predicted scene coordinates $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$0, the objective becomes

$$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$1

so $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$2 is a prior over the predicted correspondences. The paper studies pointwise Laplace priors over depth $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$3, a Wasserstein prior over the batch depth distribution, and a learned diffusion prior over the point cloud $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$4. On 7Scenes depth evaluation for ACE, Abs Rel improves from $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$5 to $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$6 with Laplace NLL and to $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$7 with the diffusion prior, while Sq Rel improves from $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$8 to $$\Omega_{1\rightarrow 2} = \{ p_1 \in \Omega_1 \mid p_2^1 \in \Omega_2 \},$$9 and $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$0, respectively; on 7Scenes relocalization, ACE improves from $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$1 to $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$2 average with diffusion, and on Stairs from $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$3 to $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$4 (Bian et al., 14 Oct 2025).

"Neural Scene Flow Prior" takes a different route: the prior is not a learned dataset prior but the architecture of a coordinate-based MLP fitted at runtime (Li et al., 2021). Scene flow is represented as

$$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$5

and optimized from random initialization on a pair of point clouds using a nearest-neighbor or Chamfer-style data term plus a backward reconstruction objective. This makes the MLP itself an implicit regularizer over continuous motion fields. The method reports on KITTI $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$6-point scene flow $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$7 and $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$8, compared with graph prior $$M_{\text{occ}}^1(p_1)=1 \quad \text{if } D_2^1(p_1) - D_2(p_2^1) > \tau.$$9 and $0.094$0, and uses the continuous field for long-term correspondence propagation by Euler integration across a sequence (Li et al., 2021).

Generative models supply another family of priors. "Correspondence Distillation from NeRF-based GAN" uses a pretrained NeRF-based GAN as a source of global structural priors through latent codes, local semantic/geometric priors through generator features, and effectively infinite synthetic NeRFs as a category-manifold prior (Lan et al., 2022). A Dual Deformation Field maps source points to a template and then to a target: $0.094$1 Feature-similarity, cycle-consistency, second-order cycle, and smoothness losses train this without correspondence labels. On keypoint transfer, the method reports [email protected] $0.094$2 and AEPE $0.094$3, compared with $0.094$4 for SIFT Flow and $0.094$5 for GLU-Net (Lan et al., 2022).

In dynamic 4D reconstruction, "Generative 4D Scene Gaussian Splatting with Object View-Synthesis Priors" shows how object-centric SDS priors can act as indirect scene correspondence priors (Chu et al., 15 Jun 2025). GenMOJO decomposes a monocular video into per-object deformable Gaussian sets, applies object-centric view-synthesis priors, and jointly renders the full scene to capture cross-object occlusions. The priors do not supervise tracks directly; they regularize the latent object geometry so that persistent correspondences survive self-occlusion, cross-object occlusion, and unseen views. On MOSE, the full model reports A-EPE $0.094$6 and M-EPE $0.094$7; removing SDS degrades these to $0.094$8 and $0.094$9, and removing joint splatting degrades A-EPE to $0.237$0 (Chu et al., 15 Jun 2025).

These methods broaden the notion of correspondence priors beyond explicit match hypotheses. Priors may act on predicted depth distributions, point-cloud plausibility, flow-field function classes, or generative canonicalizations, while still improving the final correspondences.

6. Emergent priors, empirical signatures, and limitations

A recent mechanistic perspective is provided by "On Geometric Understanding and Learned Data Priors in VGGT" (Bratulić et al., 12 Dec 2025). The paper asks whether a large feed-forward 3D transformer implements something analogous to a classical correspondence-and-geometry pipeline or relies mainly on learned appearance priors. Probing intermediate camera tokens with an MLP to predict the fundamental matrix $0.237$1 shows a sharp transition around layer $0.237$2, after which root Sampson error drops substantially and the smallest singular value of predicted $0.237$3 becomes approximately $0.237$4 smaller than the largest singular value. Attention analysis shows that cross-view top-1 matching emerges most strongly in global attention layers $0.237$5–$0.237$6, and head knockout in those layers causes major degradation in Sampson distance, up to complete failure. Under occlusion, roughly $0.237$7 of matching heads remain active for masked patches and Sampson error degrades only slightly, which the authors interpret as the model recovering or hallucinating correspondences from learned priors rather than only direct evidence (Bratulić et al., 12 Dec 2025).

Across the literature, several empirical signatures recur. Priors help most when image evidence is intrinsically ambiguous: weakly textured slanted surfaces in stereo, sparse overlap and large viewpoint gaps in RGB-D object pose, novel or unseen targets in semantic navigation, underconstrained geometry in SCR, and heavy occlusion in monocular 4D scene reconstruction (Scharstein et al., 2017, Chen et al., 20 Jun 2026, Yang et al., 2018, Bian et al., 14 Oct 2025, Chu et al., 15 Jun 2025). The quantitative pattern is consistent with that interpretation: orientation priors reduce stereo errors most on difficult Middlebury pairs, CVSI gains increase with viewpoint gap, graph priors help most on novel targets, diffusion priors help more on larger harder Indoor6 scenes than on easier ScanNet setups, and object view-synthesis priors help most on MOSE, where occlusion is frequent and prolonged (Scharstein et al., 2017, Chen et al., 20 Jun 2026, Yang et al., 2018, Bian et al., 14 Oct 2025, Chu et al., 15 Jun 2025).

The same papers also delimit the topic’s failure modes. Near-zero overlap, highly incomplete masks, or poor VFM features can defeat object-centric cross-view priors (Chen et al., 20 Jun 2026). Geometric priors in change detection depend on the quality of GFM-estimated depth and camera pose, and significant illumination changes can break geometry estimation sufficiently that Retinex or color transfer is needed as preprocessing (Liu et al., 14 Sep 2025). Depth or diffusion priors in SCR are scene-distribution dependent and strongest indoors; the learned diffusion prior is coarse and adds mapping cost even though it has no test-time relocalization overhead (Bian et al., 14 Oct 2025). Runtime neural flow priors remain vulnerable to nearest-neighbor failure under occlusion or missing parts (Li et al., 2021). Emergent priors in feed-forward 3D transformers are powerful but can hallucinate unjustified structure under asymmetric appearance shifts or genuinely underdetermined ambiguity (Bratulić et al., 12 Dec 2025).

Scene correspondence priors therefore denote a family of inductive biases rather than a single algorithmic object. What unifies them is not the specific representation—graph, disparity offset, token attention, diffusion score, morphable template, or object-centric Gaussian prior—but the role they play: they inject regularity before correspondence ambiguity becomes irrecoverable. In current work, the most effective formulations are hybrid. They combine semantic conditioning with structural preservation, or geometric validity with visibility reasoning, or canonical identity with generative completion. This suggests that the topic is best understood not as a dichotomy between geometry and priors, but as the study of how prior structure can be made usable for downstream matching, registration, parsing, tracking, and pose recovery.