Linear algebra and projective geometry
Camera poses, rays, depth, and point transforms require explicit frames.
3D and spatial vision
Spatial results are meaningful only when cameras, coordinate frames, scale, alignment, visibility, and evaluation geometry are declared.
36 maintained papers · literature reviewed through 2 October 2026
Spatial vision estimates a scene under a coordinate system, observation model, and visibility pattern. Classical reconstruction, learned depth, neural rendering, point-map regression, and feed-forward geometry differ in what is observed, what is optimized, and whether scale and coordinates are identifiable.
Prerequisite contract
Image formation and dense prediction
Photometric and feature correspondences inherit image ambiguities.
How the field developed
2006–2017
Multi-view geometry, online SLAM, and hierarchical point sets
Photo Tourism and COLMAP recover cameras and sparse/dense structure offline; ORB-SLAM adds real-time tracking, local mapping, relocalization, and loop correction; PointNet learns over unordered points, and PointNet++ adds local metric neighborhoods and hierarchical set abstraction.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 01.01 Photo Tourism: Exploring Photo Collections in 3D Photo Tourism turns a disorderly collection of photographs of the same place into a spatial browsing system. It detects and matches local features, incrementally estimates camera intrinsics and poses together with sparse 3D points, and then uses those camera–point correspondences to place each original photograph in a common coordinate frame. The contribution is an end-to-end composition: reconstruction is not the final product, but the geometric index that enables smooth photo-to-photo transitions, object-centered browsing, map alignment, historical-photo registration, and transfer of annotations between views. — 01.02 ORB-SLAM: A Versatile and Accurate Monocular SLAM System ORB-SLAM is a real-time monocular visual SLAM system organized around selected keyframes, ORB features, and a shared graph-and-point map. A tracking thread estimates the current camera pose, local mapping grows and refines the map, and loop closing recognizes a previously visited place, estimates a Sim(3) correction, and optimizes an Essential Graph before optional global bundle adjustment. The same ORB representation supports tracking, mapping, relocalization, and place recognition, making the system coherent rather than a collection of unrelated feature pipelines. Official project page · Official implementation · arXiv record 01.03 Structure-from-Motion Revisited Structure-from-Motion Revisited turns incremental SfM from a familiar recipe into a more robust general-purpose system. COLMAP augments pairwise geometry, chooses the next camera using a spatially distributed correspondence score, builds tracks with recursive robust triangulation, repeatedly alternates bundle adjustment, retriangulation, and filtering to repair drift, and groups redundant views to reduce optimization cost. The paper's lesson is that reconstruction quality comes from coordinated decisions across the full pipeline rather than from one isolated estimator. — 01.04 PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation PointNet processes a point cloud as the set it is: the same multilayer perceptron maps every point independently, a channel-wise maximum aggregates all point features into one permutation-invariant global descriptor, and task heads predict a shape class or concatenate that global vector back to local point features for per-point labels. Input and feature transformation networks learn approximate canonicalization. The architecture avoids voxel quantization and multiview rendering while providing one shared blueprint for classification, part segmentation, and scene segmentation. — 01.05 PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space PointNet++ builds a hierarchy directly on an unordered metric point set. Each set-abstraction layer samples centroids, groups neighboring points around each centroid, normalizes them into local coordinates, and applies a shared PointNet to produce a higher-level point feature. Repeating this process grows receptive fields from local geometry to global shape. Feature-propagation layers interpolate coarse features back to original points for dense segmentation. Official implementation Work: Calibrate and reconstruct one scene, contrast offline reconstruction with online SLAM, then compare global point pooling with multiscale local grouping.
Exit evidence: Transform every pose, ray, depth, and point into one declared frame; state how loops and neighborhoods are selected.
2020–2024
Neural fields, splats, and monocular depth
NeRF and Gaussian splatting optimize scene representations from views; DPT and Depth Anything learn dense relative or metric depth priors.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 02.01 NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis NeRF represents one static scene in the weights of a multilayer perceptron queried by continuous 3D position and viewing direction. The network predicts volume density from position alone and emitted RGB color from position plus direction. A differentiable volume-rendering equation composites samples along every camera ray, so calibrated photographs supervise the field without 3D ground truth. Sinusoidal positional encoding restores high spatial frequencies, while a coarse network builds a ray-wise sampling distribution that concentrates fine-network evaluations near visible content. — 02.02 Vision Transformers for Dense Prediction DPT makes an image-classification Vision Transformer usable for dense prediction. It taps token sequences from several encoder depths, resolves the ungrounded readout token, reshapes patch tokens back into image-like maps, projects and resamples them to four scales, then progressively fuses and upsamples those maps through convolutional refinement blocks. Because the transformer preserves a relatively high token resolution and every layer has a global receptive field, the decoder receives both detailed and globally coherent features for monocular depth or semantic segmentation. — 02.03 3D Gaussian Splatting for Real-Time Radiance Field Rendering 3D Gaussian Splatting replaces a repeatedly queried neural radiance-field network with an explicit, editable set of anisotropic 3D Gaussians. Each primitive carries a 3D mean, opacity, full covariance, and view-dependent spherical-harmonic color; differentiable splatting projects and alpha-composites the visible primitives. Interleaving gradient optimization with cloning, splitting, and pruning concentrates capacity where image error demands it, enabling the paper's combination of high-quality novel-view synthesis, tens-of-minutes training, and real-time 1080p rendering. — 02.04 Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data Depth Anything scales monocular relative-depth learning by combining 1.5 million labeled images with more than 62 million unlabeled images. A teacher supplies pseudo-depth for unlabeled data; strong color perturbation and CutMix make those examples harder, and a feature-alignment term keeps the depth encoder close to frozen DINOv2 semantics. The result is a general-purpose relative-depth representation evaluated zero-shot and transferred to segmentation. Official code Work: Compare reconstruction and monocular prediction under held-out view, depth, scale, and runtime metrics.
Exit evidence: Separate novel-view fidelity, geometry accuracy, and speed.
2024–2026
Feed-forward matching and multi-view geometry
DUSt3R regresses point maps; MASt3R adds geometry-grounded local matching; VGGT and OmniVGGT broaden prediction toward cameras, tracks, and scene state with fewer per-scene optimization steps.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 03.01 DUSt3R: Geometric 3D Vision Made Easy DUSt3R reframes two-view geometry as direct regression of two dense, pixel-aligned point maps in one camera's coordinate frame. A Siamese image encoder and two communicating decoders take an uncalibrated, unposed image pair and predict a 3D point plus confidence for every valid pixel in both views. Once correspondences and geometry share the same representation, depth, matching, focal length, relative pose, and multi-view reconstruction can be recovered with simple geometric procedures instead of a separately engineered feature-match-triangulate pipeline. — 03.02 Grounding Image Matching in 3D with MASt3R MASt3R treats image correspondence as a shared-3D-scene problem rather than only comparing 2D appearance. It starts from DUSt3R's paired-image Transformer, which predicts both images' dense pointmaps in one camera frame, and adds a normalized local-descriptor head trained explicitly for pixel matching. The same cross-view representation can therefore supply geometric point matches or learned feature matches. An iterative reciprocal-nearest-neighbor algorithm and a coarse-to-fine crop schedule then make dense high-resolution matching practical without discarding the 3D grounding. Official code · Official project article 03.03 VGGT: Visual Geometry Grounded Transformer VGGT is a roughly 1.2-billion-parameter transformer that turns one or many images of the same scene into cameras, depth maps, 3D point maps, and tracking features in one feed-forward pass. Its key architectural move is Alternating Attention: each block performs frame-wise self-attention to refine within-image structure and global self-attention to exchange evidence across every input view. Joint multi-task supervision makes the shared tokens a general visual-geometry representation; optional bundle adjustment can refine the already usable feed-forward prediction. — 03.04 OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer OmniVGGT extends VGGT so any subset of the scene's images may carry optional depth, intrinsics, or camera poses at training and inference. A lightweight GeoAdapter normalizes and tokenizes those geometric observations, injects camera cues through zero-initialized residual branches at every Alternating-Attention block, and adds depth cues to spatial tokens. Stochastic modality assignment teaches one model to operate with RGB alone, partial geometry, or both. The output contract remains complete cameras, depth, point maps, and confidence for all views, so sparse trusted measurements can guide the rest of a reconstruction. — Work: Evaluate unordered views, varying overlap, dynamic content, calibration errors, and long sequences.
Exit evidence: Report match accuracy, coordinate alignment, and failure recovery, not only average geometry error.
2026
Single-view proxies and interacting geometry
MILO uses a large reconstruction model as a scaffold before human/object separation and fitting; SiZeUp optimizes building heights from ordinal depth and aerial footprints, trading geometric breadth for speed.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 04.01 Reconstructing Humans and Objects in Interaction using Large Reconstruction ModelsECCV 2026 poster presentation is now verified against the official conference inventory (poster 3826), superseding the earlier author-only acceptance evidence. A Springer version of record was not verified on 2026-09-12; the existing review remains tied to its declared manuscript revision. Core mechanism: joint human-object mesh reconstruction. MILO uses a large reconstruction model to obtain a joint scene mesh from interaction images, separates human and object regions, fits an SMPL-H body to the human geometry, and optionally replaces the object region with a category/template mesh. This reading boundary matters: the pipeline reconstructs interacting humans and objects under the inputs and priors it can parse; it is not guaranteed metric truth, camera alignment, physical contact validity, or multi-object completeness. Official project · Official implementation · Official ECCV 2026 poster 04.02 SiZeUp: Fast 3D Proxy from Aerial Images via Depth Ordinal LossThe arXiv manuscript reports SIGGRAPH Asia 2026 acceptance. No final ACM proceedings record or official implementation was available by 2026-08-30. Core mechanism: footprint extrusion optimized by ordinal depth. SiZeUp converts a known building footprint into a vertical proxy, renders candidate heights into aerial views, and optimizes height parameters so the rendered depth ordering agrees with ordinal depth predicted from selected input images. This reading boundary matters: the method rapidly estimates simple building proxies from aerial evidence; it is not full photogrammetry and cannot represent arbitrary stepped, slanted, multilayer, or nonvertical structures. — Work: Evaluate single-view human–object reconstruction and ordinal-depth building-height estimation with declared camera frames, scale anchors, visibility, proxy priors, fitting objectives, and uncertainty.
Exit evidence: Report reconstruction, contact or interaction, height, speed, and failure metrics by occlusion, viewpoint, object/building type, and reference-data quality; never infer metric scale from ordinal depth alone.
2025
Residual transfer from privileged 3D teachers
MonoTAKD isolates LiDAR-exclusive BEV information after a camera assistant; CleanPose regularizes a point-based pose model with a frozen multimodal 3D teacher and causal adjustment.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 05.01 MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object DetectionPublished in the CVPR 2025 main conference proceedings, pages 22266–22275. MonoTAKD inserts a camera teaching assistant between a LiDAR BEV teacher and a monocular camera student. The assistant uses ground-truth depth during training to define an attainable camera-space BEV; cross-modal residual distillation then transfers only the LiDAR-teacher information not already represented by that assistant. Official implementation and models 05.02 CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge DistillationPublished in the ICCV 2025 main conference proceedings, pages 5990–6000. CleanPose improves category-level 6D pose estimation with two separable ideas: a causal front-door module intended to reduce spurious point-feature relations, and residual feature distillation from a frozen ULIP-2 PointBERT teacher. The zero-initialized residual adapter preserves the student's base feature while learning a teacher-aligned correction. Official implementation Work: List every training-only sensor, depth label, point representation, coordinate frame, and adapter path.
Exit evidence: Evaluate real-domain pose/detection, sensor dropout, calibration error, architecture mismatch, and measured inference cost.
2025–2026
Optical observability, phase recovery and calibrated deblurring
Compare three ways of making an inverse problem identifiable: expose latent focus texture through bandwidth/exposure control, measure phase differences on a short-path graph, or capture complementary focal planes and fit spatially varying blur. In each case, the sensing operator matters before the reconstruction prior.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 06.01 A second-order theory of texture for depth from focusAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Author spelling follows the conference PDF. The official project still lists code and data as forthcoming. A surface can be textureless in its average reflectance yet have measurable image texture from unresolved microscopic roughness. This second-order texture is subjective speckle: intensity fluctuations around the ensemble mean. Narrowband filtering reduces spectral averaging and can make that texture strong enough for passive depth from focus, provided exposure and illumination yield sufficient signal-to-noise ratio. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record Work: Which measurement creates the missing evidence, and which ambiguity remains after reconstruction?
Exit evidence: Do not turn a recoverability proxy into ground-truth depth accuracy, relative phase into absolute phase, or multi-exposure broadband capture into snapshot video. Calibration, light throughput, scene motion and operator assumptions remain part of the method.
2025–2026
Physics-guided residual refinement of frozen geometry models
Poppy uses test-time image offsets, delayed normal offsets and radiance decomposition to align a frozen RGB geometric prior with polarimetric measurements. This offers a concrete residual-correction mechanism to contrast with teacher-supervised residual KD, while preserving their different supervision and cost.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 07.01 Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal EstimationAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the PDF; the program calls the task “Surface Normal Estimation”. Training-free means no backbone retraining, not no test-time optimization. Poppy corrects a frozen RGB normal estimator with polarization measurements at test time. It learns image offsets for scene-wide corrections, normal offsets for fine detail, and a specular-radiance map that explains mixed reflection. This is physics-guided residual optimization around a pretrained model, not teacher–student knowledge distillation. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed) Work: Is the residual supervised by another model, a physical observation or task labels, and what actually changes at test time?
Exit evidence: Training-free backbone adaptation still incurs repeated backpropagation and needs polarization measurements; orthographic assumptions and sensor noise can reverse local-refinement gains.
2025–2026
Calibrated metric pose and directional consensus under uncertainty
PRISM-VO combines calibrated plenoptic depth with temporal photometric residuals; LAYS separates yaw using column-to-line voting over uncertain positions. Compare metric observability, correspondence geometry and uncertainty handling instead of treating all pose estimation as one task.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 08.01 PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle AdjustmentAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. The official project announces forthcoming code; a runnable public release was not verified. PRISM-VO couples temporal photometric tracking with metric depth from a focused plenoptic camera. Its sparse sliding-window bundle adjustment jointly estimates camera poses, inverse point depths and brightness parameters, using uncertainty-weighted inverse virtual-depth residuals. The physical micro-lens geometry makes metric scale observable; it is not recovered from a learned monocular prior. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record 08.02 Cross-View Yaw Estimation in Location Uncertainty with Line-Aligning Yaw ScoringAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Exact KITTI measurements follow Table 2, which differs from the surrounding prose; that discrepancy is retained in the review. LAYS estimates camera yaw by matching ground-image columns to aerial-image pixels and accumulating line-consensus votes. Moving a candidate camera position along the same directed radial ray preserves the implied yaw, so orientation can remain identifiable while location is uncertain. The method still searches a 3D position–yaw tensor; it does not eliminate position candidates from the computation. Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record Work: What anchors metric scale or orientation, and is translation estimated precisely or deliberately left ambiguous?
Exit evidence: Approximate pinhole baselines and differing yaw/noise protocols require explicit qualifiers. Cross-view inference is not cross-validation or cross-resolution distillation; a measurement residual is not a residual teacher.
2025–2026
Physics-informed optical fields and focus-stack geometry
Event-domain refractive inversion estimates evolving air density, while calibrated focus-stack attention estimates scene depth. Both exploit a known optical observation process; compare acquisition metadata, scene-specific optimization and synthetic-to-real validation before treating them as generic image encoders.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 09.01 Event-based Sparse-view Background-Oriented Schlieren TomographyECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #377. Publisher proceedings/version of record was not verified on 2026-09-12. EventBOS reconstructs a continuous four-dimensional air-density field directly from event streams. A coordinate MLP predicts density, velocity and pressure; differentiable refractive ray tracing turns density into changes in a textured background, which are matched to measured events. Fluid equations regularize the otherwise ambiguous one- or two-view inverse problem. ECCV Spotlight program · Conference manuscript 09.02 Zero-shot Depth from DefocusECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12. FOSSA learns metric depth from a calibrated focus stack using attention across focus settings. It alternates ordinary image-token processing with per-location stack attention conditioned on focus distance, then pools the stack and predicts dense depth. The accompanying ZEDD benchmark tests real-world zero-shot transfer rather than within-dataset fitting. ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link) Work: Which unknown is identifiable from the actual measurements, and what boundary, calibration or scene-static assumptions supply missing information?
Exit evidence: High event acquisition rate is not real-time optimization; multiple focus images provide different evidence from one monocular image.
2025–2026
Implicit-field tokenization of biomedical graphs
VesselTok compresses pre-extracted centerline graphs through pseudo-radius occupancy reconstruction, then uses latent generative models for synthesis and repair. This connects shape representation, VAE bottlenecks and graph topology without claiming to infer vessels directly from raw CT.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 10.01 VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and GenerationECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10927. Publisher proceedings/version of record was not verified on 2026-09-12. VesselTok compresses complex biomedical centerline graphs into continuous latent tokens by reconstructing a pseudo-radius occupancy field. Centerline-conditioned attention preserves thin branching structure more efficiently than densely sampling its surface. The same latent supports graph generation and conditional repair, but it does not directly encode true vessel thickness. ECCV Spotlight program · Conference manuscript · Official implementation Work: What geometric and topological information survives token compression, and which attributes were never represented?
Exit evidence: Pseudo radius is not measured vessel thickness; high overlap can coexist with worse cycle topology, and token capacity can merge or omit small structures.
2025–2026
Boundary tokens and camera-aware relative position
PuzLM converts regular puzzle borders into discrete sequences for a language model, whereas URoPE transports calibrated camera geometry into relative positional features. Both change the representation interface, but neither is a general claim that language models discover geometry without suitable inputs.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 11.01 PuzLM: Solving Jigsaw Puzzles with Sequence-to-Sequence Language ModelsECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #3026 checked on 2026-09-12. Publisher version of record was not independently verified. PuzLM turns square jigsaw assembly into discrete sequence transduction. A lightweight PCA and k-means tokenizer summarizes each piece's border; a standard encoder-decoder predicts the grid position of each piece. The solver receives symbols rather than original pixels, but those symbols still originate from visual measurements and are not evidence-free reasoning. ECCV Spotlight program · Conference manuscript 11.02 URoPE: Universal Relative Position Embedding across Geometric SpacesECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #11480 checked on 2026-09-12. Publisher version of record was not independently verified. URoPE extends rotary position embedding to tokens from different cameras or geometric spaces. A key image patch is lifted along its camera ray at fixed candidate depths, then projected into the query camera. Ordinary two-dimensional RoPE operates on the resulting positions. Geometry changes the positional interface without adding a learned depth predictor. ECCV Spotlight program · Conference manuscript · Official project or implementation Work: What geometric information is preserved in the interface, and which calibration or discretization assumptions make that possible?
Exit evidence: Square-piece puzzle reconstruction is not arbitrary fracture recovery. Parameter-free positional encoding still needs trained host models and known camera parameters.
2025–2026
Counterexamples for spatial, semantic and action grounding
Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 12.01 Why Far Looks Up: Probing Spatial Representation in Vision-Language ModelsECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #10791 checked on 2026-09-12. Publisher version of record was not independently verified. The project page describes an oral; the catalog follows the official Spotlight record. Why Far Looks Up diagnoses vertical-distance entanglement: a VLM may treat higher image position as evidence of greater depth. Controlled behavioral tests are paired with hidden-state contrastive probes, separating a correct answer from the representation that supports it. This is a diagnostic study and benchmark, not a training algorithm that guarantees three-dimensional reasoning. ECCV Spotlight program · Conference manuscript · Official project or implementation Work: Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Exit evidence: A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations.
2025–2026
Recovering, steering and testing visual concepts
Intermediate features can expose information hidden by the final representation. Compare frozen-backbone negation adaptation, language-directed representations, circuit interventions and geometry-subspace probes; distinguish information accessibility from causal control.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 13.01 Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace InterventionECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7101. Publisher proceedings/version of record was not verified on 2026-09-12. This study asks how geometry is encoded, not simply which backbone has the best depth score. It decomposes a trained linear probe by SVD, projects frozen ViT features into selected subspaces, and reuses the same frozen head. Matched aligned, random and residual projections test which directions carry the geometric information that this particular linear decoder can read. ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript Work: Which layer, token, subspace or circuit changes, and does the intervention preserve unrelated capabilities?
Exit evidence: A probe may measure accessibility rather than semantic understanding; intervention and natural distribution shift are different tests.
2025–2026
Perceptual experts, geometry and explicit grounding tokens
Continuous expert-supervised visual tokens, geometry-aware fusion and coarse-to-fine grounding alter how a language model accesses visual evidence. Compare the representation interface, auxiliary supervision and deployed computation before comparing answer scores.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 14.01 Make Geometry Matter for Spatial ReasoningECCV 2026 Oral verified in the official program (event 6086); scientific claims use the conference-hosted manuscript #482. Publisher proceedings/version of record was not verified on 2026-09-12. GeoSR asks whether a spatial-reasoning VLM actually uses its geometry tokens. It combines training-time masking of appearance tokens with learned token-and-channel-wise fusion, so geometric evidence becomes useful rather than an ignored auxiliary stream. The contribution is controlled utilization of pretrained geometry, not simply adding another encoder. ECCV program · Conference manuscript · Author project · Code Work: Does a claimed reasoning gain arise from more visual evidence, better alignment, a changed output space or a larger inference budget?
Exit evidence: Expert distillation is not proof of faithful reasoning; stronger pointing or spatial QA need not improve every tracking metric.
2025–2026
Volume tokens, persistent geometry and registration
Compare learned volume representations, streaming geometric context, flow-based rigid registration and dense image matching. Each chooses a different correspondence unit and computation path; learned priors complement rather than abolish geometric constraints.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 15.01 Volume Transformer: Revisiting Vanilla Transformers for 3D Scene UnderstandingAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. Volt makes a plain global-attention Transformer useful for 3D scene segmentation by turning occupied cubic volumes into tokens. Its central lesson is that tokenization and training data determine whether a generic architecture succeeds: a convolutional teacher supplies a useful inductive bias even when that teacher is less accurate than the trained student. Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project 15.02 Geometric Context Transformer for Streaming 3D ReconstructionAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv/released system uses the LingBot-Map name and a different author list; authors and quantitative claims here follow the official ECCV manuscript. GCT is a causal 3D reconstruction Transformer whose memory is organized by geometric function: fixed anchors establish a reference frame, recent images supply dense local overlap, and compact trajectory tokens preserve historical context. This separates what must remain detailed from what can be compressed, rather than treating every past frame identically. Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code (LingBot-Map) 15.03 Register Any Point: Scaling 3D Point Cloud Registration by Flow MatchingAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv preprint first appeared in 2025; the catalog year follows its verified 2026 conference acceptance. RAP generates the registered configuration of several point clouds directly with conditional flow matching, then recovers rigid poses from that configuration. Its crucial addition is test-time rigidity enforcement: each integration step projects the predicted clean points back to a rigid transform of each input view, reducing the mismatch between unconstrained point motion and a rigid registration task. Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project 15.04 RoMa v2: Harder Better Faster Denser Feature MatchingAccepted to ECCV 2026; spotlight presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The first preprint is from 2025; 2026 reflects the verified conference. Author order and table values follow the ECCV-hosted manuscript, not the earlier repository BibTeX. RoMa v2 combines frozen DINOv3 features with a cross-view coarse matcher and separately trained subpixel refiners. It aims to retain foundation-feature generalization while improving difficult correspondences, throughput, and uncertainty estimation. The output is a dense warp with overlap confidence and predictive covariance, not a camera pose or a generated 3D scene by itself. Official ECCV spotlight presentation · Official ECCV paper record · arXiv record · Official code Work: Which coordinates, features or correspondences are predicted, and how are rigidity, memory growth and distribution shift tested?
Exit evidence: Compact per-frame context still accumulates; multi-step integration is not one-pass inference, and improved average matching can hide benchmark regressions.
2025–2026
Generative scene representations and coordinate privacy
Autoregressive Gaussian scene generation and inversion of scene-coordinate regressors expose different consequences of learned 3D representations: what a model can generate, and what its outputs reveal about a private scene.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 16.01 GaussianGPT: Towards Autoregressive 3D Gaussian Scene GenerationAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. GaussianGPT treats a 3D Gaussian scene as a sequence of discrete spatial decisions. A sparse autoencoder compresses Gaussian attributes into quantized voxel features, and a causal Transformer alternately predicts the next occupied position and its feature code. Prefix conditioning then supports generation, completion, and outpainting with the same generative model. Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project 16.02 Seeing Through the Weights: Privacy Leakage in Scene Coordinate RegressionAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The official repository was a README with a forthcoming-code notice when checked; a runnable implementation was not verified. A scene stored only in neural-network weights can still leak its geometry. This paper shows that querying scene-coordinate regression models with unrelated proxy images produces coordinate samples that accumulate into recognizable scene structure; access to descriptors or gradients improves filtering and refinement. It challenges privacy by representation, not the cryptographic security of a protected model. Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official project · Official code repository (release announced, not yet implemented) Work: Which view, geometry or coordinate information is accessible, and do appearance quality and geometric fidelity agree?
Exit evidence: Image FID does not certify 3D coverage; coordinate-output access is stronger than a pose-only service interface.
2026
Cross-mode distillation into explicit 3D representations
FlashWorld combines a multi-view teacher, rendered-Gaussian student and learned fake-score network. Study distribution matching, consistency regularization and when OOD batches must omit adversarial targets.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 17.01 FlashWorld: High-quality 3D Scene Generation within SecondsOfficial ICLR 2026 oral and full proceedings manuscript checked. A frozen multi-view teacher guides a few-step rendered-Gaussian student using learned fake scores, consistency and OOD co-training. Shared-scene rendering is not guaranteed correct geometry. Preserve different benchmark timing protocols, proprietary-data and training-code access limitations. No WSI or cross-magnification evidence. FlashWorld distills a high-quality multi-view diffusion mode into a few-step generator that directly produces a 3D Gaussian scene. Its central idea is cross-mode distribution matching: transfer appearance quality from the multi-view teacher while rendering the student’s shared 3D representation during denoising. Official proceedings record · OpenReview · Official implementation Work: What does changing the student representation preserve, and which visual failures escape the numerical metrics?
Exit evidence: A shared rendered scene does not certify correct geometry, fewer parameters or cross-magnification transfer; preserve hardware and benchmark protocol differences.
Research branches
Multi-view reconstruction
Which cameras, matches, and scene parameters are jointly optimized?
Photo Tourism: Exploring Photo Collections in 3D · Structure-from-Motion Revisited · NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis · 3D Gaussian Splatting for Real-Time Radiance Field Rendering
Depth, point hierarchies, and grounded matches
Are neighborhoods, depth scale, and local matches geometrically consistent under density and calibration changes?
PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation · PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space · Vision Transformers for Dense Prediction · Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data · DUSt3R: Geometric 3D Vision Made Easy · Grounding Image Matching in 3D with MASt3R
General scene state
Which static and dynamic variables survive sparse or unordered observations?
VGGT: Visual Geometry Grounded Transformer · OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer
Single-view proxy reconstruction and scale-constrained geometry
Which geometry is observed, which is supplied by a pretrained proxy or footprint, and what makes scale identifiable?
Reconstructing Humans and Objects in Interaction using Large Reconstruction Models · SiZeUp: Fast 3D Proxy from Aerial Images via Depth Ordinal Loss
Training-only geometric privilege
Which spatial cue is unavailable to the deployed camera or point model, and how is only that cue isolated?
MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection · CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
Optical observability, phase recovery and calibrated deblurring
Which measurement creates the missing evidence, and which ambiguity remains after reconstruction?
Physics-guided residual refinement of frozen geometry models
Is the residual supervised by another model, a physical observation or task labels, and what actually changes at test time?
Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Calibrated metric pose and directional consensus under uncertainty
What anchors metric scale or orientation, and is translation estimated precisely or deliberately left ambiguous?
PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle Adjustment · Cross-View Yaw Estimation in Location Uncertainty with Line-Aligning Yaw Scoring
Physics-informed optical fields and focus-stack geometry
Which unknown is identifiable from the actual measurements, and what boundary, calibration or scene-static assumptions supply missing information?
Event-based Sparse-view Background-Oriented Schlieren Tomography · Zero-shot Depth from Defocus
Implicit-field tokenization of biomedical graphs
What geometric and topological information survives token compression, and which attributes were never represented?
Boundary tokens and camera-aware relative position
What geometric information is preserved in the interface, and which calibration or discretization assumptions make that possible?
PuzLM: Solving Jigsaw Puzzles with Sequence-to-Sequence Language Models · URoPE: Universal Relative Position Embedding across Geometric Spaces
Counterexamples for spatial, semantic and action grounding
Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Recovering, steering and testing visual concepts
Which layer, token, subspace or circuit changes, and does the intervention preserve unrelated capabilities?
Perceptual experts, geometry and explicit grounding tokens
Does a claimed reasoning gain arise from more visual evidence, better alignment, a changed output space or a larger inference budget?
Volume tokens, persistent geometry and registration
Which coordinates, features or correspondences are predicted, and how are rigidity, memory growth and distribution shift tested?
Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding · Geometric Context Transformer for Streaming 3D Reconstruction · Register Any Point: Scaling 3D Point Cloud Registration by Flow Matching · RoMa v2: Harder Better Faster Denser Feature Matching
Generative scene representations and coordinate privacy
Which view, geometry or coordinate information is accessible, and do appearance quality and geometric fidelity agree?
GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation · Seeing Through the Weights: Privacy Leakage in Scene Coordinate Regression
Cross-mode distillation into explicit 3D representations
What does changing the student representation preserve, and which visual failures escape the numerical metrics?
Comparison table
| Family or stage | Scientific distinction | Main risk | Anchors |
|---|---|---|---|
| Geometric optimization | Explicit cameras, correspondences, triangulation, and bundle adjustment. | Textureless, repeated, or moving regions break correspondence assumptions. | 2016 |
| Neural rendering | Optimizes a scene representation through differentiable image formation. | View fidelity can hide poor or non-identifiable geometry. | 2020 · 2023 |
| Feed-forward prior | Amortizes local point structure, matches, depth, cameras, or tracks across training scenes. | Training priors can dominate under unusual density, optics, or geometry. | 2017 · 2024 · 2024 · 2025 |
| Single-view proxies and interacting geometry | MILO scaffolds interacting human/object geometry with a large reconstruction model before separation and fitting; SiZeUp combines ordinal depth with aerial footprints to optimize building height efficiently. | Proxy hallucination, occlusion, coordinate misalignment, or an unreported scale prior can produce attractive geometry without correct metric structure. | 2026 · 2026 |
| Residual transfer from privileged 3D teachers | Separates camera-attainable structure from LiDAR-exclusive residuals and frozen point-language priors. | Privileged sensors, ground-truth depth, or approximate causal assumptions can dominate the apparent transfer gain. | 2025 · 2025 |
| Optical observability, phase recovery and calibrated deblurring | Compare three ways of making an inverse problem identifiable: expose latent focus texture through bandwidth/exposure control, measure phase differences on a short-path graph, or capture complementary focal planes and fit spatially varying blur. In each case, the sensing operator matters before the reconstruction prior. | Do not turn a recoverability proxy into ground-truth depth accuracy, relative phase into absolute phase, or multi-exposure broadband capture into snapshot video. Calibration, light throughput, scene motion and operator assumptions remain part of the method. | 2026 |
| Physics-guided residual refinement of frozen geometry models | Poppy uses test-time image offsets, delayed normal offsets and radiance decomposition to align a frozen RGB geometric prior with polarimetric measurements. This offers a concrete residual-correction mechanism to contrast with teacher-supervised residual KD, while preserving their different supervision and cost. | Training-free backbone adaptation still incurs repeated backpropagation and needs polarization measurements; orthographic assumptions and sensor noise can reverse local-refinement gains. | 2026 |
| Calibrated metric pose and directional consensus under uncertainty | PRISM-VO combines calibrated plenoptic depth with temporal photometric residuals; LAYS separates yaw using column-to-line voting over uncertain positions. Compare metric observability, correspondence geometry and uncertainty handling instead of treating all pose estimation as one task. | Approximate pinhole baselines and differing yaw/noise protocols require explicit qualifiers. Cross-view inference is not cross-validation or cross-resolution distillation; a measurement residual is not a residual teacher. | 2026 · 2026 |
| Physics-informed optical fields and focus-stack geometry | Event-domain refractive inversion estimates evolving air density, while calibrated focus-stack attention estimates scene depth. Both exploit a known optical observation process; compare acquisition metadata, scene-specific optimization and synthetic-to-real validation before treating them as generic image encoders. | High event acquisition rate is not real-time optimization; multiple focus images provide different evidence from one monocular image. | 2026 · 2026 |
| Implicit-field tokenization of biomedical graphs | VesselTok compresses pre-extracted centerline graphs through pseudo-radius occupancy reconstruction, then uses latent generative models for synthesis and repair. This connects shape representation, VAE bottlenecks and graph topology without claiming to infer vessels directly from raw CT. | Pseudo radius is not measured vessel thickness; high overlap can coexist with worse cycle topology, and token capacity can merge or omit small structures. | 2026 |
| Boundary tokens and camera-aware relative position | PuzLM converts regular puzzle borders into discrete sequences for a language model, whereas URoPE transports calibrated camera geometry into relative positional features. Both change the representation interface, but neither is a general claim that language models discover geometry without suitable inputs. | Square-piece puzzle reconstruction is not arbitrary fracture recovery. Parameter-free positional encoding still needs trained host models and known camera parameters. | 2026 · 2026 |
| Counterexamples for spatial, semantic and action grounding | Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable. | A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations. | 2026 |
| Recovering, steering and testing visual concepts | Intermediate features can expose information hidden by the final representation. Compare frozen-backbone negation adaptation, language-directed representations, circuit interventions and geometry-subspace probes; distinguish information accessibility from causal control. | A probe may measure accessibility rather than semantic understanding; intervention and natural distribution shift are different tests. | 2026 |
| Perceptual experts, geometry and explicit grounding tokens | Continuous expert-supervised visual tokens, geometry-aware fusion and coarse-to-fine grounding alter how a language model accesses visual evidence. Compare the representation interface, auxiliary supervision and deployed computation before comparing answer scores. | Expert distillation is not proof of faithful reasoning; stronger pointing or spatial QA need not improve every tracking metric. | 2026 |
| Volume tokens, persistent geometry and registration | Compare learned volume representations, streaming geometric context, flow-based rigid registration and dense image matching. Each chooses a different correspondence unit and computation path; learned priors complement rather than abolish geometric constraints. | Compact per-frame context still accumulates; multi-step integration is not one-pass inference, and improved average matching can hide benchmark regressions. | 2026 · 2026 · 2026 · 2026 |
| Generative scene representations and coordinate privacy | Autoregressive Gaussian scene generation and inversion of scene-coordinate regressors expose different consequences of learned 3D representations: what a model can generate, and what its outputs reveal about a private scene. | Image FID does not certify 3D coverage; coordinate-output access is stronger than a pose-only service interface. | 2026 · 2026 |
| Cross-mode distillation into explicit 3D representations | FlashWorld combines a multi-view teacher, rendered-Gaussian student and learned fake-score network. Study distribution matching, consistency regularization and when OOD batches must omit adversarial targets. | A shared rendered scene does not certify correct geometry, fewer parameters or cross-magnification transfer; preserve hardware and benchmark protocol differences. | 2026 |
Concept lineage
PointNet: Deep Learning on Point Sets for 3D Classification and SegmentationPointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
Global permutation-invariant point processing is extended with sampled neighborhoods and hierarchical local features.
Structure-from-Motion RevisitedNeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Recovered cameras and geometry initialize a differentiable view-synthesis representation.
DUSt3R: Geometric 3D Vision Made EasyGrounding Image Matching in 3D with MASt3R
Point-map regression is extended with a local feature head and matching objective grounded in the shared 3D representation.
DUSt3R: Geometric 3D Vision Made EasyVGGT: Visual Geometry Grounded Transformer
Pairwise point-map regression expands toward a unified feed-forward scene representation across many views.
Failure modes to audit
- Mixing camera conventions or coordinate handedness.
- Reporting aligned depth without the alignment rule.
- Using test-view overlap that leaks near-duplicates.
- Equating rendering quality with geometric accuracy.
- Ignoring dynamic objects and rolling-shutter effects.
Questions for every paper
- What coordinate frame and gauge are used?
- Which quantities are observed or predicted?
- How are scale and alignment determined?
- What happens when overlap or calibration fails?
Complete maintained paper route
Section titled “Complete maintained paper route”36 canonical readings · 2006–2026· route reviewed through October 2, 2026.
This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.
Show all 36 readings in maintained order
Photo Tourism: Exploring Photo Collections in 3D
Noah Snavely, Steven M. Seitz, and Richard Szeliski
Metadata checked
ORB-SLAM: A Versatile and Accurate Monocular SLAM System
Raúl Mur-Artal, J. M. M. Montiel, and Juan D. Tardós
Official artifacts: Official project page · Official implementation · arXiv record
Metadata checked
Structure-from-Motion Revisited
Johannes L. Schönberger and Jan-Michael Frahm
Metadata checked
PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas Guibas
Official artifacts: Official implementation
Metadata checked
PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Charles R. Qi et al.
Metadata checked
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Ben Mildenhall et al.
Metadata checked
Vision Transformers for Dense Prediction
René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun
Metadata checked
3D Gaussian Splatting for Real-Time Radiance Field Rendering
Bernhard Kerbl et al.
Metadata checked
Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data
Lihe Yang et al.
Official artifacts: Official code
Metadata checked
Grounding Image Matching in 3D with MASt3R
Vincent Leroy, Yohann Cabon, and Jérôme Revaud
Official artifacts: Official code · Official project article
Metadata checked
DUSt3R: Geometric 3D Vision Made Easy
Shuzhe Wang et al.
Metadata checked
MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D Object Detection
Hou-I Liu, Christine Wu, Jen-Hao Cheng, Wenhao Chai, Shian-Yun Wang, Gaowen Liu, Hugo Latapie, Jhih-Ciang Wu, Jenq-Neng Hwang, Hong-Han Shuai, and Wen-Huang Cheng
Published in the CVPR 2025 main conference proceedings, pages 22266–22275.
Official artifacts: Official implementation and models
Metadata checked
CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
Xiao Lin, Yun Peng, Liuyi Wang, Xianyou Zhong, Minghao Zhu, Yi Feng, Jingwei Yang, Chengju Liu, and Qijun Chen
Published in the ICCV 2025 main conference proceedings, pages 5990–6000.
Official artifacts: Official implementation
Metadata checked
VGGT: Visual Geometry Grounded Transformer
Jianyuan Wang et al.
Metadata checked
Reconstructing Humans and Objects in Interaction using Large Reconstruction Models
Agniv Chatterjee and Georgios Pavlakos
ECCV 2026 poster presentation is now verified against the official conference inventory (poster 3826), superseding the earlier author-only acceptance evidence. A Springer version of record was not verified on 2026-09-12; the existing review remains tied to its declared manuscript revision.
Official artifacts: Official project · Official implementation · Official ECCV 2026 poster
Metadata checked
SiZeUp: Fast 3D Proxy from Aerial Images via Depth Ordinal Loss
Wenjun Zhou, Yunshan Li, Qiaoyu Zhu, Weidan Xiong, Hao Zhang, Daniel Cohen-Or, and Hui Huang
The arXiv manuscript reports SIGGRAPH Asia 2026 acceptance. No final ACM proceedings record or official implementation was available by 2026-08-30.
Metadata checked
Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention
Weichen Zhou, Yawen Zou, Chunzhi Gu, Ran Dong, Haoran Xie, and Chao Zhang
ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7101. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript
Metadata checked
Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding
Kadir Yilmaz, Adrian Kruse, Tristan Höfer, Daan de Geus, and Bastian Leibe
Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project
Metadata checked
Geometric Context Transformer for Streaming 3D Reconstruction
Lin-Zhuo Chen, Jian Gao, Shangzhan Zhang, Yihang Chen, Nan Xue, Jianyuan Wang, Christian Rupprecht, Xun Cao, Xing Zhu, Yujun Shen, Yao Yao, and Yinghao Xu
Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv/released system uses the LingBot-Map name and a different author list; authors and quantitative claims here follow the official ECCV manuscript.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code (LingBot-Map)
Metadata checked
Register Any Point: Scaling 3D Point Cloud Registration by Flow Matching
Yue Pan, Tao Sun, Liyuan Zhu, Lucas Nunes, Iro Armeni, Jens Behley, and Cyrill Stachniss
Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv preprint first appeared in 2025; the catalog year follows its verified 2026 conference acceptance.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project
Metadata checked
GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation
Nicolas von Lützow, Barbara Rössle, Katharina Schmid, and Matthias Nießner
Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project
Metadata checked
Seeing Through the Weights: Privacy Leakage in Scene Coordinate Regression
Oleksii Nasypanyi, Jaemin Cho, Utku Ozbulak, Byungkon Kang, and Francois Rameau
Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The official repository was a README with a forthcoming-code notice when checked; a runnable implementation was not verified.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official project · Official code repository (release announced, not yet implemented)
Metadata checked
RoMa v2: Harder Better Faster Denser Feature Matching
Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman, Jonathan Astermark, Anders Heyden, Viktor Larsson, Mårten Wadenbäck, Michael Felsberg, and Fredrik Kahl
Accepted to ECCV 2026; spotlight presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The first preprint is from 2025; 2026 reflects the verified conference. Author order and table values follow the ECCV-hosted manuscript, not the earlier repository BibTeX.
Official artifacts: Official ECCV spotlight presentation · Official ECCV paper record · arXiv record · Official code
Metadata checked
Make Geometry Matter for Spatial Reasoning
Shihua Zhang, Qiuhong Shen, Shizun Wang, Tianbo Pan, and Xinchao Wang
ECCV 2026 Oral verified in the official program (event 6086); scientific claims use the conference-hosted manuscript #482. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV program · Conference manuscript · Author project · Code
Metadata checked
A second-order theory of texture for depth from focus
Sreekar Sai Ranganathan and Ioannis Gkioulekas
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Author spelling follows the conference PDF. The official project still lists code and data as forthcoming.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Metadata checked
PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle Adjustment
Aymeric Fleith, Julian Zirbel, Daniel Cremers, and Niclas Zeller
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. The official project announces forthcoming code; a runnable public release was not verified.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Metadata checked
Cross-View Yaw Estimation in Location Uncertainty with Line-Aligning Yaw Scoring
Taeho Kang, Nairan Zhang, Yelin Kim, Yujiao Shi, and Youngki Lee
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Exact KITTI measurements follow Table 2, which differs from the surrounding prose; that discrepancy is retained in the review.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Metadata checked
Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Irene Kim, Sai Tanmay Reddy Chakkera, Alexandros Graikos, Dimitris Samaras, and Akshat Dave
Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Title follows the PDF; the program calls the task “Surface Normal Estimation”. Training-free means no backbone retraining, not no test-time optimization.
Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record · Official code (inspected, not executed)
Metadata checked
Event-based Sparse-view Background-Oriented Schlieren Tomography
Xinyu Zhou, Shihao Hu, Peiqi Duan, Chao Xu, and Boxin Shi
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #377. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript
Metadata checked
Zero-shot Depth from Defocus
Yiming Zuo, Hongyu Wen, Venkat Subramanian, Patrick Chen, Karhan Kayan, Mario Bijelic, Felix Heide, and Jia Deng
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)
Metadata checked
VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation
Chinmay Prabhakar, Bastian Wittmann, Tamaz Amiranashvili, Paul Büschl, Ezequiel De la Rosa, Julian McGinnis, Benedikt Wiestler, Bjoern Menze, and Suprosanna Shit
ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10927. Publisher proceedings/version of record was not verified on 2026-09-12.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation
Metadata checked
PuzLM: Solving Jigsaw Puzzles with Sequence-to-Sequence Language Models
Gur Elkin, Ofir Itzhak Shahar, and Ohad Ben-Shahar
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #3026 checked on 2026-09-12. Publisher version of record was not independently verified.
Official artifacts: ECCV Spotlight program · Conference manuscript
Metadata checked
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, Yu Su, Jonathan Tremblay, Chan Hee Song, and Jaesik Park
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #10791 checked on 2026-09-12. Publisher version of record was not independently verified. The project page describes an oral; the catalog follows the official Spotlight record.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked
URoPE: Universal Relative Position Embedding across Geometric Spaces
Yichen Xie, Depu Meng, Chensheng Peng, Yihan Hu, Quentin Herau, Masayoshi Tomizuka, and Wei Zhan
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #11480 checked on 2026-09-12. Publisher version of record was not independently verified.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked
FlashWorld: High-quality 3D Scene Generation within Seconds
Xinyang Li, Tengfei Wang, Zixiao Gu, Shengchuan Zhang, Chunchao Guo, Liujuan Cao
Official ICLR 2026 oral and full proceedings manuscript checked. A frozen multi-view teacher guides a few-step rendered-Gaussian student using learned fake scores, consistency and OOD co-training. Shared-scene rendering is not guaranteed correct geometry. Preserve different benchmark timing protocols, proprietary-data and training-code access limitations. No WSI or cross-magnification evidence.
Official artifacts: Official proceedings record · OpenReview · Official implementation
Metadata checked
OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer
Hengchao Peng et al.
Metadata checked