mlsanityData to 7 Oct 2026

Emergent Extreme-View Geometry in 3D Foundation Models

arXiv 2511.22686 ↗First posted 27 Nov 2025cs.CVcited by 5

Not on the map. Why?

Abstract (arXiv)

3D foundation models (3DFMs) have recently transformed 3D vision, enabling joint prediction of depths, poses, and point maps directly from images. Yet their ability to reason under extreme, non-overlapping views remains largely unexplored. In this work, we study their internal representations and find that 3DFMs exhibit an emergent understanding of extreme-view geometry, despite never being trained for such conditions. To further enhance these capabilities, we introduce a lightweight alignment scheme that refines their internal 3D representation by tuning only a small subset of backbone bias terms, leaving all decoder heads frozen. This targeted adaptation substantially improves relative pose estimation under extreme viewpoints without degrading per-image depth or point quality. Additionally, we contribute MegaUnScene, a new benchmark of Internet scenes unseen by existing 3DFMs, with dedicated test splits for both relative pose estimation and dense 3D reconstruction. All code and data will be released.

Key facts. Cited by 5 papers in the map (data through 7 Oct 2026). Builds on 45 of its 68 references that are in the map. First posted on arXiv 27 Nov 2025.

Builds on

45 of 68 references.

References · 45

  1. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,084
  2. 2021SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformerscited by 2,923
  3. 2016Pyramid Scene Parsing Networkcited by 2,472
  4. 2023DUSt3R: Geometric 3D Vision Made Easycited by 1,526
  5. 2025VGGT: Visual Geometry Grounded Transformercited by 1,475
  6. 2019SuperGlue: Learning Feature Matching with Graph Neural Networkscited by 1,437
  7. 2017SuperPoint: Self-Supervised Interest Point Detection and Descriptioncited by 1,313
  8. 2021BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-modelscited by 1,234
  9. 2021LoFTR: Detector-Free Local Feature Matching with Transformerscited by 974
  10. 2024Grounding Image Matching in 3D with MASt3Rcited by 960
  11. 2018MegaDepth: Learning Single-View Depth Prediction from Internet Photoscited by 731
  12. 2025Continuous 3D Perception Model with Persistent Statecited by 527
33 more
  1. 2024MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motioncited by 431
  2. 2025Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Passcited by 356
  3. 2021Neural RGB-D Surface Reconstructioncited by 312
  4. 2023RoMa: Robust Dense Feature Matchingcited by 296
  5. 2019ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residualscited by 180
  6. 2025FLARE: Feed-forward Geometry, Appearance and Camera Estimation from Uncalibrated Sparse Viewscited by 176
  7. 2023PF-LRM: Pose-Free Large Reconstruction Model for Joint Pose and Shape Predictioncited by 165
  8. 2024MASt3R-SfM: a Fully-Integrated Solution for Unconstrained Structure-from-Motioncited by 161
  9. 2023PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustmentcited by 146
  10. 2024Cameras as Rays: Pose Estimation via Ray Diffusioncited by 129
  11. 2022RelPose: Predicting Probabilistic Relative Rotation for Single Objects in the Wildcited by 102
  12. 2025FastVGGT: Training-Free Acceleration of Visual Geometry Transformercited by 93
  13. 2025Easi3R: Estimating Disentangled Motion from DUSt3R Without Trainingcited by 88
  14. 2023RelPose++: Recovering 6D Poses from Sparse-view Observationscited by 87
  15. 2024What Matters in Transformers? Not All Attention is Neededcited by 83
  16. 2024MegaScenes: Scene-Level View Synthesis at Scalecited by 58
  17. 2022SparsePose: Sparse-View Camera Pose Regression and Refinementcited by 53
  18. 2021Wide-Baseline Relative Camera Pose Estimation with Directional Learningcited by 49
  19. 2021Extreme Rotation Estimation using Dense Correlation Volumescited by 42
  20. 2025AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesiscited by 39
  21. 2022The 8-Point Algorithm as an Inductive Bias for Relative Pose Prediction by ViTscited by 38
  22. 2024Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similaritycited by 36
  23. 2021Leveraging redundancy in attention with Reuse Transformerscited by 33
  24. 2025MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Secondcited by 31
  25. 2023POPE: 6-DoF Promptable Pose Estimation of Any Object, in Any Scene, with One Referencecited by 30
  26. 2018Extreme Relative Pose Estimation for RGB-D Scans via Scene Completioncited by 22
  27. 2024Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Featurescited by 20
  28. 2019Extreme Relative Pose Network under Hybrid Representationscited by 19
  29. 2022FvOR: Robust Joint Shape and Pose Optimization for Few-view Object Reconstructioncited by 19
  30. 2021Towers of Babel: Combining Images, Language, and 3D Geometry for Learning Multimodal Visioncited by 18
  31. 2024Can Generative Video Models Help Pose Estimation?cited by 7
  32. 2024Extreme Rotation Estimation in the Wildcited by 6
  33. 2025PoseCrafter: Extreme Pose Estimation with Hybrid Video Synthesiscited by 1

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 5.

5 papers.

  1. 2026Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localizationcited by 1
  2. 2026Zero-Shot Novel Depth Synthesis Using 3D Foundation Models Scene Representationscited by 0
  3. 2026RoMa-Omega: What Feed-Forward 3D Models Know About Image Matchingcited by 0
  4. 2026GeoCond: A Conditioning-Aware Reliability Adapter for Feed-Forward 3D Reconstructioncited by 0
  5. 2026When Wider Views Fail: Stress-Testing Feed-Forward 3D Reconstructioncited by 0