Target Audience#

• Readers interested in generating point clouds or constructing 3D models from multiple images.
• Readers interested in feature matching between two images, as well as change detection using images or point clouds.
• Readers who want to understand the relationship between traditional technologies such as SfM, MVS, and SLAM and newer 3D vision models such as DUSt3R and VGGT.

Summary#

• DUSt3R represents a turning point from solving 3D geometry through a sequence of explicit geometric operations to directly predicting 3D geometry using learned neural networks.
• Its central innovation is the pointmap, which directly represents the 3D coordinates corresponding to each image pixel. This makes it possible to address 3D reconstruction, depth estimation, image matching, and camera pose estimation within a unified framework.
• DUSt3R's ideas have evolved into MASt3R, Fast3R, MUSt3R, CUT3R, and VGGT, with applications in SLAM, Gaussian Splatting, and robotics.
• However, traditional geometric techniques such as SfM and Bundle Adjustment have not become obsolete. Combining learned 3D predictions with classical geometric optimization remains an important direction.

1. Introduction: What Did DUSt3R Change?#

"DUSt3R: Geometric 3D Vision Made Easy" is a paper that marked an important turning point in the history of 3D vision research.[1]

In short, DUSt3R bridges the transition from solving 3D geometry through explicit geometric procedures to having neural networks trained on large-scale 3D data directly infer geometry.

Traditionally, reconstructing 3D models or point clouds from multiple photographs required several processing stages, including feature matching, camera pose estimation, and triangulation.

DUSt3R proposed a different approach: instead of executing these processes individually, it directly predicts 3D spatial coordinates from images.

This idea has subsequently influenced research in 3D reconstruction, feature matching, SLAM, and 3D Gaussian Splatting.

A 2026 survey also describes the family of methods originating from DUSt3R as part of the research areas known as "feed-forward 3D reconstruction" and "geometric foundation models."[2]

This article traces the development from earlier 3D vision research to MASt3R and VGGT, explaining where DUSt3R fits into the evolution of modern 3D vision technology.

2. 3D Vision Before DUSt3R#

2.1 The Classical 3D Reconstruction Pipeline#

Traditionally, reconstructing 3D geometry from multiple photographs generally involves the following pipeline:

Images
   ↓
Feature Extraction
   ↓
Image Matching
   ↓
Fundamental / Essential Matrix
   ↓
Camera Pose Estimation
   ↓
Triangulation
   ↓
Bundle Adjustment
   ↓
Multi-View Stereo
   ↓
3D Point Cloud

COLMAP is a representative implementation of this approach.[3]

COLMAP first estimates camera poses and sparse 3D point clouds through Structure-from-Motion (SfM). It then generates dense 3D point clouds using Multi-View Stereo (MVS) techniques, such as PatchMatch Stereo.

This approach is strongly grounded in geometric theory.

Its major components include:

• Camera intrinsic parameters (K)
• Camera pose (R,t)
• Epipolar geometry
• Triangulation
• Bundle Adjustment

These techniques remain extremely important today.

2.2 Weaknesses of Classical Methods#

However, classical reconstruction pipelines also have limitations.

One particularly important limitation is that errors in pixel correspondences can propagate through subsequent stages of the pipeline.

For example, image matching becomes difficult under conditions such as:

• Low-texture regions
• Repetitive patterns
• Large viewpoint changes
• Motion blur

If feature correspondences cannot be estimated correctly, subsequent camera pose estimation and triangulation may also be affected.

Starting in the late 2010s, researchers increasingly explored replacing individual components of classical pipelines with neural networks.

3. Replacing Geometric Processing with Deep Learning#

3.1 MVSNet: Learning Depth Estimation#

MVSNet (2018) is a representative study that introduced deep learning into Multi-View Stereo.[4]

MVSNet uses known camera parameters to construct a cost volume and estimates depth using a neural network.

However, camera calibration and camera poses are still required as inputs.

In other words, MVSNet does not eliminate the traditional geometric framework. Instead, it replaces part of the pipeline with a learned method.

3.2 SuperGlue and LoFTR: Learning Feature Matching#

Deep learning also made significant progress in feature matching.

SuperGlue (2020) introduced a learnable feature matching mechanism based on Graph Neural Networks and optimal transport.[5]

LoFTR (2021) uses Transformer-based cross-attention to estimate correspondences between images.[6]

In particular, LoFTR can estimate dense correspondences even in low-texture regions where traditional keypoint detectors often struggle.

The general philosophy during this period can be summarized as follows:

Preserve the classical 3D reconstruction pipeline while improving individual components using deep learning.

For example, even when feature matching is replaced by LoFTR, the overall structure often still includes Fundamental Matrix estimation and camera pose estimation afterward.

4. CroCo: Teaching Transformers Cross-View Geometry#

An important predecessor to DUSt3R is CroCo and its successor, CroCo v2.[7]

CroCo introduces a pretraining method called Cross-view Completion.

The idea is to use images of the same scene captured from different viewpoints and train a network to reconstruct masked image regions.

In other words, the masked-image reconstruction task is extended to multiple views.

CroCo v2 demonstrated that large-scale Transformer pretraining could achieve strong performance in stereo matching and optical flow.

An important result was that high performance could be achieved without relying heavily on combinations of conventional mechanisms such as cost volumes, warping, and iterative refinement.

The CroCo v2 paper describes this direction as a step toward a "universal vision model."

The key insight is that Transformers can learn not only semantic image information but also geometric relationships between different viewpoints.

DUSt3R takes this idea one step further.

5. The Core Idea of DUSt3R: Pointmaps#

5.1 Eliminating the Need to Estimate Cameras First#

The most important idea behind DUSt3R is to directly predict 3D coordinates without first estimating camera parameters.[1]

In conventional approaches, processing generally follows this sequence:

[ \text{correspondence} \rightarrow K,R,t \rightarrow \text{triangulation} \rightarrow XYZ ]

First, image correspondences are estimated. Then camera intrinsic parameters and poses are recovered, and triangulation is used to compute 3D coordinates.

DUSt3R instead performs the following operation:

[ (I_1,I_2) \overset{\mathrm{Transformer}}{\longrightarrow} (X_1,X_2) ]

Here, (I_1,I_2) are the input images, and (X_1,X_2) are the predicted pointmaps.

5.2 What Is a Pointmap?#

A conventional depth map estimates a single depth value for each image pixel:

[ (u,v)\rightarrow z ]

A pointmap, on the other hand, directly estimates three-dimensional coordinates for each pixel:

[ (u,v)\rightarrow (X,Y,Z) ]

In other words, it directly represents where each image pixel is located in 3D space.

Furthermore, DUSt3R predicts the 3D points associated with two images in a shared coordinate system.

This represents a major change.

With conventional depth maps, integrating depth information from different cameras requires not only depth values but also camera intrinsic parameters and camera poses.

Depth Map
    +
Camera Intrinsics
    +
Camera Pose
    ↓
3D Coordinate Conversion
    and Integration

With a pointmap, each pixel is already associated with a 3D coordinate.

Image
  ↓
DUSt3R
  ↓
Pointmap
  ↓
3D Coordinates

This makes it possible to obtain information required for 3D reconstruction more directly.

5.3 Unifying Multiple 3D Vision Tasks#

By using pointmaps, DUSt3R can address multiple problems that were traditionally treated separately.

Examples include:

• 3D reconstruction
• Depth estimation
• Pixel correspondence
• Relative camera pose estimation
• Absolute camera pose estimation
• Focal length estimation

Traditionally, many of these problems required separate estimation algorithms or processing steps.

With DUSt3R, the necessary information can be derived from the predicted pointmaps.

Therefore, the innovation of DUSt3R is not simply that it uses Transformers.

Its deeper contribution is changing the fundamental output representation of 3D vision from camera parameters, depth, and matching to pointmaps.

6. DUSt3R's Position in the History of 3D Vision#

The evolution so far can be summarized as follows:

Period	Representative Research / Technology	Core Idea
Classical era	SIFT / SfM / COLMAP / MVS	Explicitly solve geometric relationships
Around 2018 onward	MVSNet	Learn parts of the geometric pipeline
2020–2021	SuperGlue / LoFTR	Learn image correspondences
2021–2023	DPT / CroCo	Learn dense geometric information using Transformers
2023–2024	DUSt3R	Directly regress pointmaps and unify geometric tasks
2024–2025	MASt3R / Fast3R / CUT3R / MUSt3R	Extend to matching, multi-view, and streaming
2025–2026	VGGT and related models	Develop foundation models for 3D geometry

Research such as DPT demonstrated that Transformers are effective for dense prediction tasks.[8]

CroCo further showed that Transformers could learn geometric relationships between multiple views.[7]

Building on these developments, DUSt3R introduced pointmaps as a direct output representation.

From this perspective, DUSt3R is not simply another depth estimation model.

It represents a change in how 3D vision problems are formulated.

7. Limitations of DUSt3R: Two-View Models and Global Alignment#

DUSt3R does not reconstruct a large 3D scene from an arbitrary number of images in a single forward pass.

The original DUSt3R is fundamentally a two-view model.[1]

For example, when processing 100 images, DUSt3R estimates pointmaps for selected image pairs and then combines them into a common coordinate system through global alignment.

Image Pairs
    ↓
DUSt3R
    ↓
Pairwise Pointmaps
    ↓
Global Alignment
    ↓
Unified 3D Scene

Mathematically, the process can be represented as:

[ \text{pairwise DUSt3R} \rightarrow \text{global alignment} \rightarrow \text{scene} ]

Consequently, computation becomes a challenge as the number of images increases.

When processing many image pairs, the number of possible pairs can grow on the order of (O(N^2)).

Addressing this limitation became an important research direction in 3D vision between 2024 and 2026.

8. MASt3R: Combining 3D Estimation and Feature Matching#

One of the most direct successors to DUSt3R is MASt3R.[9]

MASt3R adds a head that predicts dense local features alongside DUSt3R's pointmaps.

Although DUSt3R is robust to large viewpoint changes, it was not necessarily optimized for precise pixel matching.

MASt3R focuses on improving this aspect.

DUSt3R can be characterized as:

[ \text{DUSt3R}

\text{3D geometry} ]

MASt3R extends this idea to:

[ \text{MASt3R}

\text{3D geometry} + \text{high-quality matching} ]

In other words, geometric 3D information and image correspondence estimation are handled within the same network.

For readers interested in feature matching between images, the evolution from DUSt3R to MASt3R is particularly important.

9. MASt3R-SLAM and SLAM3R: Using 3D Predictions for SLAM#

MASt3R's approach has also been applied to SLAM.

9.1 MASt3R-SLAM#

MASt3R-SLAM, introduced in 2025, is a real-time dense SLAM method that uses MASt3R as a source of 3D reconstruction priors.[10]

Traditional SLAM systems generally combine feature extraction, tracking, and camera pose estimation.

MASt3R-SLAM instead adopts a structure similar to the following:

RGB Images
    ↓
3D Priors from MASt3R
    ↓
Tracking / Fusion / Loop Closure
    ↓
Dense SLAM

The paper reports dense monocular SLAM at approximately 15 FPS.

9.2 SLAM3R#

SLAM3R follows a similar direction.[11]

It directly regresses pointmaps from RGB video and incrementally integrates local 3D information, reporting dense reconstruction at more than 20 FPS.

The important change is in the system design.

Traditional SLAM systems typically extract image features and use them to estimate camera motion.

In contrast, DUSt3R-based approaches make it possible to predict 3D information from images first and then use that information for tracking.

This is an example of how DUSt3R's pointmap representation has influenced real-time 3D perception.

10. Fast3R and MUSt3R: Reconstructing Multiple Images Together#

Another major limitation of DUSt3R is its dependence on pairwise inference and global alignment.

Fast3R and MUSt3R address this issue.

10.1 Fast3R#

Fast3R is positioned as a multi-view generalization of DUSt3R.[12]

It takes multiple images as input to a Transformer and produces pointmaps in a common coordinate system within a single forward pass.

[ (I_1,I_2,\ldots,I_N) \rightarrow (X_1,X_2,\ldots,X_N) ]

DUSt3R requires a process similar to:

Process Image Pairs
        ↓
Pairwise 3D Estimation
        ↓
Global Alignment
        ↓
Unified 3D Model

Fast3R moves toward:

Multiple Images
       ↓
Transformer
       ↓
Pointmaps in a Shared
Coordinate System

This makes multi-image 3D reconstruction more direct.

10.2 MUSt3R#

MUSt3R is another extension developed by the authors of DUSt3R.[13]

It moves away from pairwise reconstruction followed by global alignment toward directly processing multiple views.

It also introduces a memory mechanism to support larger image collections and online processing.

The difference can be summarized as follows.

DUSt3R

Estimate 3D geometry from pairs of images and integrate the results afterward.

Fast3R / MUSt3R

Estimate 3D geometry from multiple images directly in a shared coordinate system.

11. CUT3R: Maintaining a Persistent 3D State#

Another important research direction following DUSt3R is CUT3R.[14]

CUT3R updates an internal state whenever a new camera image arrives.

Conceptually, the process can be written as:

[ I_1 \rightarrow S_1 ]

[ I_2 + S_1 \rightarrow S_2 ]

[ I_3 + S_2 \rightarrow S_3 ]

Here, (S_i) represents an internal state that preserves information from previously observed images.

The model continuously updates 3D information while outputting pointmaps in a shared coordinate system.

This approach is particularly suitable for robotic perception.

For a robot, 3D perception is not necessarily a task of processing hundreds of photographs all at once.

Instead, it is often a continuous process in which new camera images arrive over time, requiring the robot to update its understanding of the surrounding 3D world.

CUT3R extends DUSt3R's pointmap regression concept toward a persistent 3D state.

12. VGGT: Toward Geometric Foundation Models#

One of the most representative developments following DUSt3R is VGGT (Visual Geometry Grounded Transformer).[15]

VGGT was developed by researchers from Meta and Oxford and received the CVPR 2025 Best Paper Award.[16]

Given one to hundreds of images, VGGT directly predicts several forms of geometric information from a single Transformer.

These include:

• Camera intrinsics
• Camera extrinsics
• Depth maps
• Point maps
• 3D point tracks

DUSt3R demonstrated the idea of unifying multiple 3D vision tasks around pointmaps.

VGGT takes this approach further by jointly predicting camera parameters, depth, pointmaps, and point tracking information.

This reflects a broader shift in thinking.

Rather than constructing separate 3D algorithms for individual tasks, the goal becomes building a large model capable of understanding 3D scene geometry itself.

Models of this type are increasingly described as Geometric Foundation Models (GFMs).[2]

From this perspective, DUSt3R was an important turning point in the development of geometric foundation models.

13. Connections to Modern 3D Model Construction#

13.1 Gaussian Splatting and NeRF#

DUSt3R-based research is also connected to 3D representation technologies such as Gaussian Splatting and NeRF.

For example, a common pipeline for 3D Gaussian Splatting is:

Multiple Images
       ↓
COLMAP
       ↓
Camera Poses + Point Cloud
       ↓
Gaussian Splatting
       ↓
3D Model

This pipeline requires camera pose estimation and point cloud initialization, often using COLMAP.

However, COLMAP-based estimation can become difficult when only a few images are available or when viewpoint differences are large.

Recent research therefore explores the use of DUSt3R, MASt3R, and VGGT.

Multiple Images
       ↓
DUSt3R / MASt3R / VGGT
       ↓
Camera Poses + Geometry
       ↓
Gaussian Splatting
       ↓
3D Model

For example, the official VGGT implementation provides functionality to export predictions in COLMAP format, making it possible to use the results with Gaussian Splatting libraries such as gsplat.[17]

InstantSplat is another representative study that uses geometric priors from the DUSt3R/MASt3R family for pose-free and sparse-view Gaussian Splatting.[18]

These are concrete examples of how DUSt3R-related models can be incorporated into practical 3D modeling workflows.

13.2 Implications for Point Cloud Generation and Feature Matching#

For point cloud generation from images, DUSt3R's pointmap representation is important because it directly provides 3D coordinates associated with image pixels.

For feature matching, models such as MASt3R extend this approach by combining 3D geometric information with dense image features.

These technologies are relevant not only to 3D model reconstruction but also to processing tasks that require determining correspondences between different images.

For image-based change detection, differences in camera viewpoints must often be considered before meaningful comparisons can be made.

This typically requires establishing image correspondences or geometric alignment.

Technologies such as DUSt3R and MASt3R are relevant because they provide ways to estimate the 3D information and image correspondences needed for such comparisons.

However, DUSt3R itself is not a model that directly performs change detection.

14. Research in 2026: Scaling, Generative Models, and Robotics#

Research building on DUSt3R and VGGT has continued to evolve in 2026.

14.1 VGG-T³: Scaling to Large Image Collections#

VGG-T³, presented at CVPR 2026, studies ways to improve the computational and memory efficiency of Transformers when handling large numbers of images.[19]

It explores feed-forward reconstruction for image collections on the scale of 1,000 images.

This extends the idea of directly predicting 3D geometry from images toward larger reconstruction tasks.

14.2 Gen3R: Combining 3D Reconstruction and Generative Models#

Gen3R combines VGGT's geometric representation with video diffusion models.[20]

Conceptually, this can be expressed as:

[ \text{geometry foundation model} + \text{generative model} ]

This direction goes beyond reconstructing 3D geometry from observed images and begins to connect 3D reconstruction with 3D generation.

14.3 Robotics and Vision-Language-Action Models#

Geometric Foundation Models are also being studied for robotics applications.

Research published in 2026 explores incorporating models such as VGGT into Vision-Language-Action (VLA) systems.[21]

The key idea is to let geometric foundation models provide accurate 3D information that Vision-Language Models (VLMs) and VLA models may struggle to infer.

However, such applications are still at the research stage.

15. Does DUSt3R Make Classical Geometry Obsolete?#

An important point is that DUSt3R does not make classical geometry unnecessary.

Its success should not be understood as simply replacing geometry with neural networks.

Instead, an increasingly important approach is to combine:

[ \boxed{\text{learned geometric prior}} + \boxed{\text{classical geometric optimization}} ]

For example, VGGT provides functionality not only to use predicted camera parameters and points directly but also to refine predictions through Bundle Adjustment.[17]

Real-world 3D reconstruction still faces several challenges, including:

• High-resolution images
• Very long image sequences
• Unknown camera parameters
• Reflective and transparent objects
• Dynamic objects
• Scenes outside the training data distribution

A 2026 review describes DUSt3R-related methods as a developing technology rather than a complete replacement for traditional SfM and MVS.[2]

Challenges remain in accuracy, scalability, dynamic scenes, and other areas.

A comparative study involving aerial photographs also reported that DUSt3R, MASt3R, and VGGT have advantages under extremely sparse-view conditions, but they do not completely replace traditional approaches in large-scale or high-resolution settings.[22]

Therefore, combining learned 3D prediction with classical geometric optimization remains an important direction in modern 3D vision.

16. The Research Lineage Centered on DUSt3R#

The research developments discussed so far can be summarized in the following diagram.

           Classical 3D Vision
                    │
      matching → pose → triangulation
                    │
          bundle adjustment → MVS
                    │
           ┌────────┴────────┐
           ↓                 ↓
    learned matching     learned depth
   SuperGlue / LoFTR     MVSNet / DPT
           │                 │
           └────────┬────────┘
                    ↓
             CroCo / CroCo v2
          cross-view Transformer
                    │
                    ↓
              ┌── DUSt3R ──┐
              │  pointmap  │
              │ regression │
              └─────┬──────┘
                    │
           ┌────────┴──────────┐
           ↓                   ↓
        MASt3R           CUT3R / Spann3R
    matching + 3D         spatial memory
           │                   │
           ↓                   ↓
     MASt3R-SLAM            online 3D
           │
           ├────── Fast3R / MUSt3R
           │         multi-view
           │
           └──────────┐
                      ↓
                  VGGT etc.
          Geometric Foundation Model
                      │
             ┌────────┼─────────┐
             ↓        ↓         ↓
            SLAM     3DGS     Robotics
                     NeRF       VLA

This diagram illustrates how DUSt3R connects classical 3D vision techniques with geometric foundation models.

Traditionally, the primary research direction was to introduce neural networks into individual tasks such as matching, depth estimation, and pose estimation.

DUSt3R changed this by first predicting a general-purpose 3D representation—the pointmap—and then deriving the necessary information from that representation.

This idea subsequently developed into research such as MASt3R, Fast3R, CUT3R, MUSt3R, and VGGT.

17. Conclusion: Where Does DUSt3R Fit?#

DUSt3R can be positioned as an important turning point in establishing the concept of foundation models for 3D vision.

Neural network-based 3D vision methods existed long before DUSt3R.

However, many earlier approaches focused on solving individual tasks such as stereo matching, depth estimation, feature matching, and camera pose estimation using neural networks.

DUSt3R introduced a different formulation.

Instead of solving each task separately, it first predicts a general-purpose 3D representation—a pointmap—and derives depth, pose, matching, and other information from it.

This marks a shift from viewing 3D vision as a collection of separate geometric algorithms toward treating it as a problem of directly inferring 3D spatial structure.

The research timeline can be summarized as follows:

• 2024: DUSt3R establishes pointmap-centered 3D prediction.
• 2024–2025: MASt3R, Fast3R, CUT3R, and MUSt3R extend the approach to matching, multi-view reconstruction, and streaming.
• 2025: VGGT advances the development of Geometric Foundation Models.
• 2026: Research expands toward large-scale 3D reconstruction, SLAM, generative models, and robotics.

Viewed in this context, DUSt3R's approach is more than a temporary research trend.

It has developed into a broader research paradigm.

A 2026 survey describes this progression under the theme "From DUSt3R to VGGT and beyond."[2]

For readers interested in point cloud generation from images, 3D model construction, feature matching, or change detection, DUSt3R is an important paper for understanding how modern 3D vision technologies have evolved.

Looking ahead, one of the central challenges will remain determining how to best combine learned geometric representations with traditional geometric optimization.

References#

[1] S. Wang et al., “DUSt3R: Geometric 3D Vision Made Easy,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. https://openaccess.thecvf.com/content/CVPR2024/html/Wang_DUSt3R_Geometric_3D_Vision_Made_Easy_CVPR_2024_paper.html

[2] “From DUSt3R to VGGT and Beyond: A Survey on Feed-Forward 3D Reconstruction,” 2026. https://www.sciencedirect.com/science/article/pii/S2096579626000203

[3] J. L. Schönberger et al., COLMAP Documentation. https://colmap.github.io/

[4] Y. Yao et al., “MVSNet: Depth Inference for Unstructured Multi-view Stereo,” Proceedings of the European Conference on Computer Vision (ECCV), 2018. https://openaccess.thecvf.com/content_ECCV_2018/html/Yao_Yao_MVSNet_Depth_Inference_ECCV_2018_paper.html

[5] P.-E. Sarlin et al., “SuperGlue: Learning Feature Matching with Graph Neural Networks,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. https://openaccess.thecvf.com/content_CVPR_2020/html/Sarlin_SuperGlue_Learning_Feature_Matching_With_Graph_Neural_Networks_CVPR_2020_paper.html

[6] J. Sun et al., “LoFTR: Detector-Free Local Feature Matching with Transformers,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. https://openaccess.thecvf.com/content/CVPR2021/html/Sun_LoFTR_Detector-Free_Local_Feature_Matching_With_Transformers_CVPR_2021_paper.html

[7] P. Weinzaepfel et al., “CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow,” Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. https://openaccess.thecvf.com/content/ICCV2023/html/Weinzaepfel_CroCo_v2_Improved_Cross-view_Completion_Pre-training_for_Stereo_Matching_and_ICCV_2023_paper.html

[8] R. Ranftl et al., “Vision Transformers for Dense Prediction,” Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. https://openaccess.thecvf.com/content/ICCV2021/html/Ranftl_Vision_Transformers_for_Dense_Prediction_ICCV_2021_paper.html

[9] V. Leroy et al., “MASt3R: Grounding Image Matching in 3D with MASt3R,” 2024. https://arxiv.org/abs/2406.09756

[10] K. Murai et al., “MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Murai_MASt3R-SLAM_Real-Time_Dense_SLAM_with_3D_Reconstruction_Priors_CVPR_2025_paper.html

[11] “SLAM3R,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://mlanthology.org/cvpr/2025/liu2025cvpr-slam3r/

[12] “Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Yang_Fast3R_Towards_3D_Reconstruction_of_1000_Images_in_One_Forward_CVPR_2025_paper.html

[13] “MUSt3R: Multi-view Network for Stereo 3D Reconstruction,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Cabon_MUSt3R_Multi-view_Network_for_Stereo_3D_Reconstruction_CVPR_2025_paper.html

[14] “Continuous 3D Perception Model with Persistent State,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Wang_Continuous_3D_Perception_Model_with_Persistent_State_CVPR_2025_paper.html

[15] “VGGT: Visual Geometry Grounded Transformer,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VGGT_Visual_Geometry_Grounded_Transformer_CVPR_2025_paper.html

[16] VGGT Project Page. https://vgg-t.github.io/

[17] Meta Research, VGGT Official Repository. https://github.com/facebookresearch/vggt

[18] NVIDIA Research, InstantSplat Official Repository. https://github.com/NVlabs/InstantSplat

[19] “VGG-T³: Offline Feed-Forward 3D Reconstruction at Scale,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. https://openaccess.thecvf.com/content/CVPR2026/html/Elflein_VGG-T3_Offline_Feed-Forward_3D_Reconstruction_at_Scale_CVPR_2026_paper.html

[20] “Gen3R: 3D Scene Generation Meets Feed-Forward Reconstruction,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. https://openaccess.thecvf.com/content/CVPR2026/html/Huang_Gen3R_3D_Scene_Generation_Meets_Feed-Forward_Reconstruction_CVPR_2026_paper.html

[21] Research on Integrating Geometric Foundation Models with Vision-Language-Action Models, arXiv, 2026. https://arxiv.org/abs/2605.24642

[22] Comparative Study of DUSt3R, MASt3R, and VGGT for Aerial Imagery, arXiv, 2025. https://arxiv.org/abs/2507.14798