The Position of VGGT in the Evolution of 3D Vision

Target Audience#

This article is intended for readers interested in:

• Generating point clouds from multiple images and constructing 3D models.
• Image feature matching, image registration, and change detection between two images.
• Understanding how recent deep learning-based 3D reconstruction methods relate to technologies such as SfM, NeRF, 3D Gaussian Splatting, and SLAM.

Summary#

• VGGT represents an important shift from traditional multi-stage geometric computation toward unified 3D geometry prediction using Transformers.
• The evolution from SfM/MVS through VGGSfM, DUSt3R, and MASt3R led to models capable of jointly estimating camera parameters, depth, 3D points, and point tracks from multiple images.
• VGGT can be integrated with existing technologies, including NeRF, 3D Gaussian Splatting, and SLAM.
• As of 2026, research is moving toward Geometry Foundation Models, with developments such as VGGT-Ω and MapAnything.
• However, VGGT does not completely replace classical geometry. Combining fast predictions from pretrained models with geometric optimization, such as Bundle Adjustment, remains an important direction.

1. Where Does VGGT Fit in the History of 3D Computer Vision?#

VGGT (Visual Geometry Grounded Transformer) is a Transformer-based model designed to estimate the 3D structure of a scene from images.

Introduced in 2025, it received the Best Paper Award at CVPR 2025. [1]

From a historical perspective, VGGT occupies an important position in the development of 3D computer vision.

In simple terms, it represents a transition:

From a world where 3D geometry is solved through feature matching, geometric computation, and iterative optimization, to a world where a Transformer trained on large-scale 3D data directly predicts scene geometry.

This is more than an improvement in 3D reconstruction accuracy.

Traditionally, reconstructing a 3D model from images required combining multiple processing stages, including feature detection, feature matching, camera pose estimation, triangulation, and optimization.

VGGT instead treats many of these tasks as related prediction problems within a shared Transformer model.

As a result, its significance extends beyond point cloud generation to image feature matching, camera pose estimation, 3D Gaussian Splatting, SLAM, and other applications.

This article examines the research developments leading to VGGT and explains how it relates to technologies used in 2026.

2. How Was 3D Reconstruction Traditionally Performed?#

2.1 3D Reconstruction Using SfM and MVS#

Two representative classical approaches to 3D reconstruction are Structure from Motion (SfM) and Multi-View Stereo (MVS).

These techniques reconstruct the 3D structure of objects and environments from multiple images while estimating the camera positions and orientations from which the images were captured.

One of the best-known implementations is COLMAP. [2]

A typical SfM/MVS pipeline follows these steps:

Images → Feature Detection → Feature Matching → Camera Pose Estimation → Triangulation → Bundle Adjustment → Dense MVS → 3D Reconstruction

The main stages are:

1. Feature Detection: Identify distinctive locations in images that can be recognized across different views.
2. Feature Matching: Establish correspondences between locations representing the same physical points in different images.
3. Camera Pose Estimation: Estimate camera positions and orientations using image correspondences.
4. Triangulation: Calculate the 3D coordinates of points observed from multiple viewpoints.
5. Bundle Adjustment (BA): Jointly optimize camera parameters and 3D point positions.
6. Dense MVS: Reconstruct a denser representation of the 3D scene.

This approach has important advantages.

Because camera models, projective geometry, and epipolar geometry are explicitly represented, the resulting 3D coordinates can be explained mathematically.

However, the pipeline can become fragile when distinctive features are difficult to detect, viewpoint differences are large, textures are limited, or surfaces are reflective.

It also requires careful integration of multiple processing stages.

2.2 Replacing Individual Stages with Deep Learning#

From the late 2010s to the early 2020s, deep learning was gradually introduced into traditional SfM pipelines.

Initially, researchers did not attempt to eliminate SfM entirely.

Instead, they replaced individual components with neural networks.

For example:

• Using neural networks for feature extraction.
• Using neural networks for feature matching.
• Using neural networks for depth estimation.

In other words, the research direction was to replace individual components of classical 3D reconstruction pipelines with learned models.

2.3 End-to-End Reconstruction with VGGSfM#

VGGSfM (CVPR 2024) advanced this approach further. [3]

It combines point tracking, camera estimation, and differentiable Bundle Adjustment into an end-to-end trainable SfM system.

This represented a substantial improvement over traditional pipelines.

However, VGGSfM still preserved the overall SfM problem formulation and retained geometric optimization through Bundle Adjustment.

Its fundamental approach was therefore to modernize the traditional SfM pipeline using deep learning.

The next major conceptual shift came with DUSt3R.

3. How DUSt3R Changed the Way We Think About 3D Reconstruction#

One of the most important predecessors of VGGT is DUSt3R (CVPR 2024). [4]

DUSt3R introduced a remarkably different approach.

Traditionally, reconstruction followed this sequence:

Estimate Cameras → Find Correspondences → Triangulate Points → Construct 3D Geometry

DUSt3R proposed a different question:

What if a Transformer could directly predict the 3D location of every pixel from just two images?

3.1 The Point Map Representation#

An important concept introduced by DUSt3R is the Point Map.

A Point Map assigns a 3D coordinate to each pixel in an image.

While a conventional Depth Map represents the depth associated with each pixel, a Point Map directly represents the corresponding 3D position.

DUSt3R predicts Point Maps from image pairs without requiring known camera poses or prior camera calibration.

From these predicted Point Maps, additional information can be derived, including depth, correspondences, focal length, and camera poses. [4]

In traditional reconstruction, camera poses are estimated before computing 3D points.

DUSt3R reverses this relationship.

By predicting 3D geometry directly, it becomes possible to derive camera information and image correspondences from the reconstructed geometry.

This represented a fundamental shift in how the 3D reconstruction problem was formulated.

3.2 Limitations of DUSt3R#

Despite its advantages, DUSt3R had an important limitation.

Its fundamental processing unit was an image pair.

For example, reconstructing a scene from 10 or 100 images requires processing numerous image pairs and subsequently aligning their Point Maps through global alignment.

As the number of images increases, pairwise processing and global integration become increasingly important challenges.

3.3 The Development of MASt3R#

MASt3R (ECCV 2024) extended the ideas behind DUSt3R to image matching. [5]

By using features grounded in 3D geometry, MASt3R improves correspondence estimation between images, including cases involving large viewpoint changes.

An important implication is that image matching and 3D reconstruction no longer need to be treated as entirely separate tasks.

Traditionally, feature matching was used as an intermediate step toward reconstructing 3D geometry.

With DUSt3R and MASt3R, learned 3D representations can also be used to improve image matching.

This reverse relationship is particularly relevant to image registration and change detection.

For example, when comparing two images captured from different camera positions, simple pixel-wise differences cannot reliably distinguish viewpoint changes from actual changes in the scene.

By considering 3D geometry when establishing correspondences, it becomes possible to develop image comparison and alignment methods that account for viewpoint differences.

However, DUSt3R and MASt3R should not be understood as models that directly perform every type of change detection.

4. What Exactly Did VGGT Change?#

VGGT extends learning-based geometric estimation, as developed in DUSt3R and MASt3R, toward unified processing of multiple images.

Given anywhere from a single image to hundreds of images, VGGT can jointly predict several types of geometric information using one neural network. [1][6]

4.1 Multiple Geometric Outputs from One Transformer#

VGGT produces four major types of outputs.

Output	Description	Related Applications
Camera Parameters	Camera intrinsic and extrinsic parameters	Camera pose estimation, 3D reconstruction
Depth Maps	Per-pixel depth information	Depth and distance estimation
Point Maps	3D coordinates corresponding to image pixels	Point cloud generation, 3D modeling
Point Tracks	Corresponding points tracked across images	Image correspondence, alignment, tracking

Traditionally, these quantities were often computed through separate algorithms.

VGGT predicts them jointly from a shared Transformer representation using multiple output heads.

Camera pose estimation, depth estimation, 3D reconstruction, and point tracking are therefore treated as related aspects of understanding a common 3D scene rather than as entirely independent tasks.

4.2 Frame-wise Attention and Global Attention#

The architecture of VGGT is also notable.

Images are converted into tokens using a DINO-based encoder, followed by alternating Frame-wise Attention and Global Attention operations. [6]

Conceptually, the architecture can be represented as follows:

Images → Encoder → Frame-wise Attention → Global Attention → Frame-wise Attention → Global Attention → … → Geometric Predictions

Frame-wise Attention processes information within individual images.

Global Attention exchanges information across multiple images.

This allows the network to learn relationships between different views of the same 3D environment.

In classical SfM, these relationships are established through explicit feature matching, epipolar geometry, triangulation, and related computations.

VGGT's Global Attention can be interpreted as implicitly learning some of the relationships that these classical operations were designed to handle.

4.3 The Idea of Over-complete Prediction#

One of VGGT's distinctive design choices is over-complete prediction.

If a Depth Map and camera pose are known, the corresponding Point Map can be computed.

Similarly, camera information can be recovered from Point Maps under appropriate conditions.

Consequently, the model's outputs contain mathematical redundancy.

In conventional algorithm design, predicting multiple representations of related information might appear inefficient.

VGGT adopts the opposite approach.

By jointly supervising camera parameters, depth, Point Maps, and Point Tracks, it encourages the network's internal representations to become grounded in 3D geometry.

In other words, the model learns not only visual features but also geometric relationships from multiple complementary perspectives.

This is central to the meaning of Visual Geometry Grounded Transformer.

5. The Evolution from SfM to VGGT#

The major research generations can be summarized as follows.

Generation	Fundamental Idea	Processing Approach
SfM / COLMAP	Solve geometry through mathematical computation	Matching → PnP → Triangulation → BA
Deep SfM / VGGSfM	Make SfM components trainable	Neural Networks + Geometric Optimization
DUSt3R	Directly predict 3D Point Maps	Image Pairs → Point Maps → Global Alignment
MASt3R	Extend 3D-aware representations to image matching	Image Pairs → 3D-aware Matching
VGGT	Jointly infer scene geometry using a Transformer	Multiple Images → Cameras + Depth + Points + Tracks
Post-VGGT Research	Develop Geometry Foundation Models	Backbones for SLAM, 3DGS, Robotics, etc.

The research direction has evolved through several major stages.

Stage 1: Reconstruct 3D geometry through explicit computation

Feature matching and projective geometry are used to estimate camera poses and 3D coordinates.

Stage 2: Introduce deep learning into individual geometric operations

Feature extraction, matching, and depth estimation are progressively replaced or enhanced by neural networks.

Stage 3: Predict 3D coordinates directly

DUSt3R introduces direct Point Map prediction from image pairs.

Stage 4: Jointly infer geometric information from multiple images

VGGT combines depth, camera pose, geometric points, and tracks within one Transformer.

Stage 5: Develop general-purpose Geometry Foundation Models

The goal expands toward reusable pretrained models that support SLAM, 3D reconstruction, neural rendering, and other tasks.

However, it is important not to interpret this history as a single linear sequence in which VGGT alone created the entire transition.

Fast3R was released on arXiv in January 2025, while VGGT appeared in March 2025. [7][1]

Other approaches, including CUT3R, emerged around the same period.

Multiple research groups were independently exploring the transition from pairwise reconstruction toward direct processing of multiple views.

VGGT is best understood as one of the representative achievements of this broader research movement.

6. VGGT Represents a Shift Toward 3D Foundation Models#

The historical significance of VGGT extends beyond improving an individual reconstruction algorithm.

In image recognition, research evolved through developments such as:

ImageNet Classifiers → ViT → DINO / CLIP → General-purpose Visual Foundation Models

Rather than constructing a dedicated model for every recognition task, researchers increasingly use general-purpose visual representations pretrained on large image datasets.

VGGT suggests that a similar transformation may be occurring in 3D geometry.

Traditionally, SfM, MVS, camera pose estimation, depth estimation, and point tracking were investigated as separate problems, each with its own specialized methods.

VGGT instead demonstrates the possibility of training a Transformer on large-scale 3D data and using its learned geometry representations across multiple tasks.

The VGGT paper also reports improvements when its pretrained features are transferred to non-rigid point tracking and feed-forward novel-view synthesis. [1][8]

A conceptual analogy can be drawn with large language models.

Instead of combining many specialized language-processing modules, modern Transformers learn broad language representations through pretraining.

Although language and 3D geometry involve different constraints and problem structures, both fields exhibit a transition:

From improving task-specific algorithms toward training large, reusable foundation models.

7. How Does VGGT Connect to Modern 3D Modeling Technologies?#

VGGT is increasingly relevant to technologies such as NeRF and 3D Gaussian Splatting.

7.1 Relationship with NeRF#

NeRF (Neural Radiance Fields) learns a representation of a 3D scene from multiple images and uses it to synthesize images from new viewpoints.

The original NeRF method assumes that camera poses are known. [9]

Therefore, a traditional workflow often requires camera pose estimation using SfM or a similar method before training NeRF.

A typical pipeline is:

Images → Camera Pose Estimation with COLMAP → NeRF → Novel-view Synthesis

VGGT provides an alternative way to estimate the geometric information required by such a pipeline.

Conceptually:

Images → VGGT → Camera Parameters + Depth + Point Cloud → NeRF

An important distinction is that VGGT and NeRF are not competing technologies.

VGGT estimates 3D geometry from images.

NeRF represents a scene and generates images from new viewpoints.

Their purposes are different, and they can complement one another.

7.2 Relationship with 3D Gaussian Splatting#

Camera poses and initial point clouds are also important in practical 3D Gaussian Splatting (3DGS) pipelines.

Traditionally, COLMAP is often used to obtain camera parameters and an initial point cloud.

With VGGT, a possible workflow becomes:

Uncalibrated Images → VGGT → Cameras + Depth + Point Cloud → 3D Gaussian Splatting

The official VGGT implementation supports exporting predictions in COLMAP format, allowing them to be used with NeRF and Gaussian Splatting libraries such as gsplat. [10]

AnySplat (SIGGRAPH Asia 2025) provides another example.

It generates 3D Gaussians from uncalibrated images using a feed-forward approach and uses pretrained VGGT as a geometry prior during training. [11]

The relationship can therefore be summarized as:

• VGGT: A model for perceiving and estimating 3D geometry.
• 3D Gaussian Splatting: A representation for rendering 3D scenes.

They are complementary technologies.

This distinction is particularly relevant for applications that construct 3D models from ordinary images.

8. Applications to SLAM and Robotics#

Another important application area is Simultaneous Localization and Mapping (SLAM).

SLAM estimates the position of a moving camera or robot while simultaneously constructing a map of its surroundings.

It is widely relevant to robotics, autonomous navigation, and augmented reality.

8.1 VGGT-SLAM#

VGGT-SLAM, developed at MIT, applies VGGT to SLAM. [12]

It divides long video sequences into smaller submaps, reconstructs each submap using VGGT, and connects the resulting reconstructions through geometric optimization.

A conceptual processing structure might look like this:

First 30 Frames → VGGT → Local 3D Map

Next 30 Frames → VGGT → Local 3D Map

Following 30 Frames → VGGT → Local 3D Map

These local maps are then aligned through a SLAM backend to construct a larger global map.

The actual number of frames and segmentation strategy depend on the system design.

The important concept is to use submaps to handle long sequences that would be difficult to process as a single input.

8.2 Development Toward VGGT-SLAM++#

VGGT-SLAM++ appeared in 2026. [13]

It combines VGGT predictions with techniques such as local Bundle Adjustment and loop closure.

This reveals an important direction in post-VGGT research.

Rather than completely replacing classical geometry with AI, researchers are increasingly combining them.

VGGT quickly produces useful initial estimates of 3D structure.

Classical geometric optimization improves consistency between estimates and helps correct errors accumulated over long sequences.

Their roles can be summarized as:

VGGT: Rapidly produce useful initial 3D estimates and geometry priors.

Classical Geometry: Improve long-term and global geometric consistency.

This hybrid approach is particularly appropriate for robotics applications involving extended motion through 3D environments.

9. Connections to Feature Matching and Image Change Detection#

For readers interested in matching features between images or detecting differences between photographs, VGGT offers an important new perspective.

However, it is necessary to distinguish capabilities demonstrated directly by the research from potential downstream applications.

9.1 Traditional Feature Matching#

In traditional feature matching, distinctive image points and their descriptors are extracted from each image.

The descriptors are compared to establish candidate correspondences.

Geometric constraints are then used to reject incorrect matches or estimate the relationship between the cameras.

The process involves finding visually similar features and determining whether the resulting correspondences are geometrically consistent.

9.2 Changes Introduced by DUSt3R, MASt3R, and VGGT#

DUSt3R directly estimates Point Maps from image pairs.

MASt3R uses geometry-grounded representations to estimate correspondences.

VGGT processes multiple images jointly and predicts geometric information that includes Point Tracks.

As a result, image correspondence estimation is moving beyond simple similarity between 2D visual descriptors.

The geometry of the underlying 3D scene becomes an integral part of how relationships between images are modeled.

9.3 Potential Applications to Change Detection#

Consider detecting changes between two photographs captured at different times or from different positions.

Simple pixel-wise differences may be affected by:

• Differences in camera position and orientation.
• Changes in the apparent shape of objects due to viewpoint.
• Different apparent displacements caused by depth.
• Regions that look different even though the underlying scene has not changed.

In these situations, estimating correspondences with an understanding of 3D geometry can improve the foundation for image registration and comparison.

VGGT provides outputs such as camera poses, Point Maps, and Point Tracks that could be used in such a processing pipeline.

For example, images might first be aligned using estimated 3D geometry, after which regions that fail to agree could be treated as candidates for change detection.

However, VGGT is not itself a dedicated change detection model.

Additional processing is required to distinguish actual structural changes from lighting differences, object motion, occlusions, and reconstruction errors.

VGGT is therefore better understood as a potential geometric foundation for aligning and comparing images, rather than a complete change detection solution.

10. From VGGT to Geometry Foundation Models#

Between 2025 and 2026, additional research began generalizing the ideas behind VGGT.

10.1 MapAnything#

MapAnything, developed by Meta and CMU, explores general-purpose Transformer models for 3D geometry. [14]

An important extension is that input information does not need to consist exclusively of images.

The model can also incorporate information such as:

• Images.
• Known camera intrinsics.
• Camera poses.
• Depth information.
• Partial 3D observations.

MapAnything is designed to reconstruct metric 3D geometry using various combinations of available information.

Its feed-forward framework supports tasks including SfM, MVS, metric monocular depth estimation, localization, and depth completion.

This extends the conceptual scope beyond an individual 3D reconstruction algorithm toward a general-purpose model that accepts and produces 3D geometric information.

10.2 VGGT-Ω#

VGGT itself evolved into VGGT-Ω (VGGT-Omega) in 2026.

The official VGGT repository describes it as "the next step of VGGT." [10]

VGGT-Ω is a CVPR 2026 research project that improves reconstruction accuracy, efficiency, and support for dynamic scenes compared with the original VGGT. [15]

The researchers also report that increasing model size and training data can produce predictable improvements in 3D reconstruction performance.

This result is significant.

Traditionally, much of 3D vision research focused on designing specialized geometric algorithms for individual problems.

VGGT-Ω represents another direction: training increasingly capable geometry models on larger 3D datasets.

The idea of scaling pretrained models, which played an important role in the development of large language models, is becoming relevant to 3D vision as well.

11. Why VGGT Has Not Completely Replaced Classical SfM#

Despite its capabilities, VGGT does not eliminate the need for classical geometric methods.

There are three major reasons.

11.1 Absolute Metric Scale#

The scale of the 3D geometry estimated by standard VGGT does not necessarily correspond directly to real-world metric units such as exactly 1.00 meter.

During training, VGGT normalizes 3D quantities based on Point Map distances. [16]

Consequently, even when the relative geometry of a reconstruction is correct, its coordinates may not directly represent real-world measurements.

Applications that require accurate physical dimensions need additional information or processing to establish metric scale.

11.2 Computational Cost with Large Numbers of Images#

VGGT uses Global Attention to integrate information across images.

Although powerful, this architecture introduces computational and GPU memory challenges as the number of images increases.

VGGT-SLAM addresses long sequences by dividing them into smaller submaps. [12]

VGG-T³, introduced in 2026, explicitly investigates the difficulty of scaling VGGT-style attention to larger numbers of views in terms of computation and memory. [17]

Thus, while VGGT advances direct multi-view processing, efficient reconstruction from very large image collections remains an active research problem.

11.3 Geometric Consistency#

VGGT can rapidly predict plausible 3D geometry using learned representations.

However, its predictions are not automatically guaranteed to satisfy every projective geometric constraint exactly.

For example, camera poses and 3D point predictions may contain inconsistencies across multiple images.

Geometric optimization techniques such as Bundle Adjustment can improve consistency by jointly refining the estimated parameters.

The official VGGT implementation supports optional Bundle Adjustment after exporting reconstruction results in COLMAP format. [10]

This illustrates an important direction in practical system design:

Estimate 3D Geometry Quickly with a Foundation Model → Improve Consistency Through Geometric Optimization When Necessary → Construct or Render 3D Models

The resulting systems combine learning-based inference with established geometric methods.

12. Reviewing the Research Evolution as a Timeline#

The major developments can be summarized chronologically.

Classical SfM / MVS#

3D structure is reconstructed through feature matching, camera geometry, and optimization.

COLMAP is a representative implementation. [2]

↓

Deep Feature Extraction / Matching / Depth#

Neural networks improve individual components of classical SfM pipelines.

↓

VGGSfM (2024)#

The SfM pipeline becomes end-to-end trainable.

However, geometric optimization through Bundle Adjustment remains part of the process. [3]

↓

DUSt3R (2024)#

Point Maps are predicted directly from image pairs.

The traditional sequence of estimating cameras before triangulating points is fundamentally reconsidered. [4]

↓

MASt3R (2024)#

3D-aware representations are used to improve image matching. [5]

↓

Fast3R / CUT3R / VGGT and Related Methods (2025)#

Research increasingly explores direct processing of multiple views rather than relying exclusively on image pairs.

VGGT jointly predicts camera parameters, depth, Point Maps, and Point Tracks. [1][7]

↓

Geometry Foundation Models (2025–2026)#

Research expands toward systems such as VGGT-SLAM, AnySplat, MapAnything, and VGGT-Ω. [11][12][14][15]

VGGT-style models become reusable geometry backbones for SLAM, 3DGS, robotics, and related tasks.

↓

Current Direction#

Feed-forward Learned Geometry + Classical Optimization + Neural Rendering

This represents a hybrid architecture in which pretrained models estimate geometry, classical algorithms improve consistency when necessary, and techniques such as NeRF and 3D Gaussian Splatting are used to represent or render reconstructed scenes.

13. The Essential Contribution of VGGT: From Multi-stage Algorithms to Pretrained Geometric Inference#

The importance of VGGT is not simply that a faster alternative to traditional SfM has appeared.

Classical reconstruction required a sequence of operations such as feature matching, camera pose estimation, triangulation, and Bundle Adjustment during inference.

VGGT instead learns many of the geometric regularities involved in these problems from large-scale 3D training data.

At inference time, a shared Transformer predicts multiple geometric quantities directly.

The central change is therefore that some of the computation and optimization traditionally performed at inference time has effectively been transferred into pretraining.

However, this does not make geometry itself unnecessary.

Predictions from pretrained models can still benefit from classical optimization, and the resulting geometry can be used with NeRF or 3D Gaussian Splatting to construct practical 3D systems.

For applications involving point cloud generation, 3D model construction, feature matching, image registration, and change detection, VGGT should not be viewed as a complete replacement for every existing processing stage.

Instead, it offers a unified way to obtain information about the geometric structure of a scene.

In this sense, VGGT is not merely an alternative to COLMAP. It introduces a new geometry backbone that can be integrated with COLMAP, SLAM, 3D Gaussian Splatting, and other technologies.

The emergence of VGGT-Ω and MapAnything further illustrates the transition from specialized reconstruction algorithms toward general-purpose Geometry Foundation Models.

From reconstructing 3D scenes through chains of specialized algorithms to directly predicting 3D geometry using pretrained models.

VGGT is one of the most important research contributions for understanding this transition.

References#

[1] Wang et al., "VGGT: Visual Geometry Grounded Transformer," CVPR, 2025.
https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VGGT_Visual_Geometry_Grounded_Transformer_CVPR_2025_paper.html

[2] COLMAP, "COLMAP Documentation," Project Documentation.
https://colmap.github.io/

[3] Wang et al., "VGGSfM: Visual Geometry Grounded Deep Structure From Motion," CVPR, 2024.
https://openaccess.thecvf.com/content/CVPR2024/html/Wang_VGGSfM_Visual_Geometry_Grounded_Deep_Structure_From_Motion_CVPR_2024_paper.html

[4] Wang et al., "DUSt3R: Geometric 3D Vision Made Easy," CVPR, 2024.
https://openaccess.thecvf.com/content/CVPR2024/html/Wang_DUSt3R_Geometric_3D_Vision_Made_Easy_CVPR_2024_paper.html

[5] Leroy et al., "MASt3R: Grounding Image Matching in 3D with MASt3R," ECCV, 2024.
https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/9080_ECCV_2024_paper.php

[6] VGGT Authors, "VGGT: Visual Geometry Grounded Transformer," Official Project Page, 2025.
https://vgg-t.github.io/

[7] Yang et al., "Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass," arXiv:2501.13928, 2025.
https://arxiv.org/abs/2501.13928

[8] Visual Geometry Group, University of Oxford, "Publications," Research Publications.
https://www.robots.ox.ac.uk/~vgg/publications/index.php?idAuthor=131&pg=2

[9] Mildenhall et al., "NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis," ECCV, 2020.
https://www.ecva.net/papers/eccv_2020/papers_ECCV/html/1473_ECCV_2020_paper.php

[10] Meta Research, "VGGT: Visual Geometry Grounded Transformer," GitHub Repository, 2025–2026.
https://github.com/facebookresearch/vggt

[11] AnySplat Authors, "AnySplat," SIGGRAPH Asia, 2025, GitHub Repository.
https://github.com/OpenRobotLab/AnySplat

[12] VGGT-SLAM Authors, "VGGT-SLAM," arXiv:2505.12549, 2025.
https://arxiv.org/abs/2505.12549

[13] Mandal et al., "VGGT-SLAM++," CVPR Workshops, 2026.
https://openaccess.thecvf.com/content/CVPR2026W/VOCVALC/html/Mandal_VGGT-SLAM_CVPRW_2026_paper.html

[14] MapAnything Authors, "MapAnything," arXiv:2509.13414, 2025.
https://arxiv.org/abs/2509.13414

[15] VGGT-Ω Authors, "VGGT-Ω," CVPR, 2026, Official Project Page.
https://vggt-omega.github.io/

[16] Wang et al., "VGGT: Visual Geometry Grounded Transformer," CVPR, 2025, Paper PDF.
https://robots.ox.ac.uk/~vgg/publications/2025/Wang25/wang25.pdf

[17] Elflein et al., "VGG-T³: Offline Feed-Forward 3D Reconstruction at Scale," CVPR, 2026.
https://openaccess.thecvf.com/content/CVPR2026/papers/Elflein_VGG-T3_Offline_Feed-Forward_3D_Reconstruction_at_Scale_CVPR_2026_paper.pdf