A three-dimensional Gaussian sputtering initialization method based on Fast3R

Abstract

The application discloses a three-dimensional Gaussian sputtering initialization method based on Fast3R, which performs Fast3R fast reasoning on an input image sequence under low resolution; calculates Anchor scores of each view based on the product combination of quality scores and mutual view coverage scores, takes Fast3R output as initialization, performs feature extraction and matching, triangulation to obtain sparse trajectory constraints, introduces dense consistency constraints based on point graph projection, constructs a joint geometric optimization model, and robustly iteratively solves camera external parameters and internal parameters. The QSP strategy is introduced, high-confidence and strong cross-view consistency point clouds are preferentially retained under uniform spatial voxel constraints, the point cloud size is controlled, and outlier and floating points are reduced, thereby providing high-integrity, high-precision and controllable-scale initialization for three-dimensional Gaussian sputtering. The sensitivity of Fast3R to the selection of the first frame is effectively reduced, the stability and point cloud quality of large-scale scene reconstruction are improved, and the 3DGS training convergence and rendering effect are improved under the premise of ensuring the initialization efficiency.

Description

Technical Field

[0001] This invention relates to the field of 3D reconstruction and graphics rendering technology, specifically to a 3D Gaussian sputtering initialization method based on Fast3R. Background Technology

[0002] With the development of novel representation methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), 3D reconstruction and new perspective synthesis based on multi-view image sequences have attracted widespread attention in fields such as digital content generation, cultural tourism digitization, industrial measurement, and robot perception. Compared with implicit representations such as NeRF, 3DGS uses explicit Gaussian points as carriers, which has the advantages of fast training speed, high rendering efficiency, and excellent detail representation. However, the rendering quality and convergence stability of 3DGS largely depend on the accuracy of camera intrinsic and extrinsic parameters and the quality of the initial point cloud provided during the initialization stage. In existing engineering workflows, the initialization of 3DGS usually relies on Structure from Motion (SfM) pipelines (such as COLMAP) to estimate camera parameters and generate sparse point clouds. However, traditional SfM generally relies on manual feature extraction and matching, which makes strong assumptions about image texture, viewpoint overlap, and scene structure.

[0003] To overcome the limitations of traditional SfM (Simulation-based Modeling), learning-based multi-view matching and geometric inference methods, such as DUST3R and MASt3R, have emerged in recent years. These methods can generate denser correspondences between views and improve pose estimation. Building on this, Fast3R further proposes a feedforward 3D reconstruction framework that can process thousands of images and output point cloud and pose information in a single forward inference, providing a new technical path for efficient reconstruction of large-scale scenes. However, when Fast3R is used directly as the initialization scheme for downstream 3DGS, there are still issues such as global coordinates and poses being sensitive to anchor views, insufficient accuracy of camera intrinsic and extrinsic parameters, a lack of robust joint optimization mechanisms for 3DGS, and a lack of quality awareness and spatial structure constraints in point cloud export, making it difficult to provide an initial point set with "controllable scale and high effective information density". Summary of the Invention

[0004] Purpose of the Invention: To address the problems mentioned in the background art, this invention discloses a 3D Gaussian sputtering initialization method based on Fast3R. It performs Fast3R fast inference on the input image sequence, calculates the Anchor score for each viewpoint based on the product of quality score and mutual view coverage score, and uses the Fast3R output as initialization. Simultaneously, it performs feature extraction and matching, triangulation to obtain sparse trajectory constraints, introduces dense consistency constraints based on point map projection, constructs a joint geometric optimization model, performs robust iterative solution for camera extrinsic and intrinsic parameters, and introduces a quality-aware spatial Poisson sampling (QSP) strategy. Under the constraint of spatial voxel uniformity, it prioritizes the retention of point clouds with high confidence and strong cross-view consistency, controls the point cloud size, and reduces outliers and floating points. It outputs camera parameters and point cloud results conforming to the COLMAP format, providing a high-completeness, high-precision, and controllable-scale initialization for 3D Gaussian sputtering.

[0005] Technical solution:

[0006] This invention discloses a three-dimensional Gaussian sputtering initialization method based on Fast3R, the method comprising the following steps:

[0007] S1 acquires the scene image sequence for 3D Gaussian sputtering initialization, preprocesses the scene image sequence, and obtains a set of input images of uniform specifications;

[0008] S2: Perform low-resolution Fast3R inference on the input image set to obtain the global point map, local point map and confidence map corresponding to each view, and construct view quality and mutual view correlation information based on the inference results;

[0009] S3: Calculate the Anchor score of each view based on view quality and mutual view correlation information, select the best Anchor view from the input image set; rearrange the image sequence with the best Anchor view as the first frame, generate an image list file consistent with the rearrangement and generate a matching constraint file;

[0010] S4: Perform high-resolution Fast3R inference on the rearranged image sequence, and use iterative PnP with RANSAC to solve the initial camera extrinsic parameters for each view;

[0011] S5: Construct a joint geometric optimization model: Using the initial camera parameters and Fast3R point map as initialization priors, and combining feature extraction and matching, sparse reprojection constraints formed by triangulation, and dense consistency constraints formed by cross-view reprojection of point map, robust joint optimization of camera extrinsic parameters is performed to obtain the camera parameter set.

[0012] S6: Construct candidate point clouds based on camera parameter sets and point map results, introduce QSP strategy, prioritize the retention of high-quality point clouds while satisfying the constraint of uniform spatial distribution, and output the initial point cloud for 3DGS training;

[0013] S7: Convert and export the camera parameter set and the initial point cloud as input for the 3DGS task.

[0014] Furthermore, the low-resolution Fast3R inference described in S2 includes:

[0015] All view images are input into the Fast3R network to obtain the local point map, global point map, confidence map, and coarse camera pose estimate for each view. The confidence map is statistically analyzed in the spatial dimension to obtain the view confidence mean, which is used as the first component of view quality. The mutual viewing intensity between views is estimated based on the point map visibility and overlapping area of ​​any two views, and a mutual viewing association table is formed accordingly.

[0016] Furthermore, the specific calculation steps for the Anchor score described in S3 are as follows:

[0017] Anchor scoring includes a view quality item and a mutual visibility item: the view quality item is obtained by weighting the view confidence statistic, depth stability and view diversity, and is used to measure the geometric reliability of each view; the mutual visibility item is obtained by the overlap ratio between the view and other views, and the higher the overlap ratio, the greater the mutual visibility; the anchor score is a combination of the quality item and the mutual visibility item according to preset weights, and the view with the highest score is selected as the best anchor view, reducing the global coordinate system bias and pose instability caused by random selection in the first frame.

[0018] Furthermore, the overlap ratio statistics of the mutual visibility term include:

[0019] Select a local point map of the view and map it to a shared coordinate system through a local-global alignment transformation. Project the mapped 3D points onto the image plane of other views. Calculate the proportion of the projection falling within the effective area of ​​the image as the view pair overlap ratio. Aggregate the overlap ratios of each view pair to obtain the mutual visibility term of the view.

[0020] Furthermore, the matching constraint file mentioned in S3 includes selecting the top-K adjacent views with the highest overlap ratio in the mutual view association table for each view to form a candidate matching pair, and writing the candidate matching pair into the Pairs constraint file to reduce invalid matches and reduce the amount of subsequent optimization calculations.

[0021] Furthermore, the specific steps of S4 include:

[0022] The initial camera parameter set is solved by focal length candidate search and RANSAC-PnP: a focal length candidate set is set for each view, and the corresponding intrinsic parameter matrix is ​​calculated for each candidate focal length; under the intrinsic parameter matrix, the correspondence between high-confidence 3D points and their 2D pixels is selected as PnP input based on the confidence map; RANSAC-PnP is used to solve the camera extrinsic parameters, and the focal length candidate with the smallest reprojection error is used as the optimal focal length and extrinsic parameter output for that view; when the image sequence comes from the same physical camera, the optimal focal length of the Anchor view is used as the globally shared focal length.

[0023] Furthermore, the construction of the sparse reprojection constraint term described in S5 includes:

[0024] Based on the Pairs images generated by S3, feature extraction and matching are performed on the files to obtain sparse corresponding points across views. The above sparse corresponding points are combined with the camera parameters estimated by Fast3R as priors to perform triangulation to obtain sparse 3D points, and a reprojection error objective function is constructed as a sparse constraint term for the joint optimization objective.

[0025] Furthermore, the construction of the dense consistency constraint term includes:

[0026] Several high-quality pixels are sampled on the point map of the source view to obtain their three-dimensional coordinates; the three-dimensional coordinates are projected onto the target view using the current camera parameters to obtain the projection position; corresponding features are sampled on the multi-scale image features of the source view and the target view respectively, feature consistency error is constructed and accumulated in a robust manner to form a dense consistency constraint term; the sparse reprojection constraint term and the dense consistency constraint term are weighted and summed as a joint optimization objective to iteratively update the camera extrinsic parameters.

[0027] Furthermore, the QSP strategy described in S6 includes uniform sampling within the view and global voxel constraints: For each view, a pixel-level quality map is constructed based on the confidence map, reprojection residual, local-global alignment residual, and gradient curvature information; the image is divided into a regular grid, and the 3D point corresponding to the pixel with the highest quality score in each grid is selected as a candidate point, and high-quality candidate points are supplemented from the entire image when there are insufficient candidate points; all view candidate points are mapped to a 3D voxel grid, and only the top-K candidate points are retained in each voxel according to their quality scores, so as to obtain a sampling point cloud with uniform spatial distribution and controllable scale.

[0028] Beneficial effects:

[0029] 1. This invention performs Fast3R coarse inference on the entire scene image under a unified low-resolution configuration to obtain the point map and confidence information of each view. It also constructs an Anchor scoring mechanism that includes "view quality—mutual view coverage / overlap—visibility constraints," automatically selecting the best Anchor and rearranging the sequence. This strategy effectively avoids the propagation of global coordinate system bias and structural errors introduced by random or low-quality first frames, improving the robustness and consistency of extrinsic parameter initialization from the source.

[0030] 2. This invention utilizes the complementary advantages of SfM and Fast3R for camera parameter solving and optimization. It leverages the sub-pixel-level camera intrinsic and extrinsic parameters and sparse trajectory observations provided by SfM as high-precision geometric anchoring, ensuring scale, drift, and intrinsic parameter stability. Simultaneously, it utilizes the dense point map and confidence scores output by Fast3R to construct co-visible adjacency relationships, providing density consistency constraints and quality weighting to supplement geometric information and suppress outlier influences in areas with weak textures, repetitive textures, or insufficient disparity. Through this complementary mechanism, robustness and global consistency in complex scenes are significantly improved while maintaining the high accuracy of SfM, better meeting the requirements of 3DGS for camera parameter consistency.

[0031] 3. This invention introduces quality perception and spatial structure constraint (QSP sampling), and constructs a quality map by integrating confidence, reprojection residual, local-global alignment residual and curvature. First, it selects the best value uniformly according to the grid in the view, and then it retains the value equally by using voxel Top-K in the global view. This significantly reduces noise, redundancy and floating points while maintaining the geometric structure and key details, and avoids the loss of details and density unevenness caused by random or threshold sampling.

[0032] 4. This invention converts the optimized camera parameters and the initial point cloud after QSP sampling into the COLMAP standard data format for output, which can be directly integrated into mainstream 3DGS rendering without additional adaptation. Attached Figure Description

[0033] Figure 1 This is a flowchart of the overall method of the present invention;

[0034] Figure 2 This is a preferred schematic diagram of the Anchor of the present invention;

[0035] Figure 3 This is a schematic diagram of the joint geometry optimization of the present invention;

[0036] Figure 4 This is a diagram of the quality-perceived spatial Poisson sampling strategy of the present invention;

[0037] Figure 5 This is a comparison image of the ablation experiment rendering effects in an embodiment of the present invention;

[0038] Figure 6This is a diagram illustrating a practical application of an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] like Figure 1 As shown, this invention discloses a three-dimensional Gaussian sputtering initialization method based on Fast3R, the method steps of which are as follows:

[0041] Step 1: Obtain the scene image sequence for 3D Gaussian sputtering (3DGS) initialization. The image sequence can be obtained from a public dataset or a multi-view image sequence obtained by taking panoramic photos around the target scene. The target scene can be an indoor space scene or an outdoor environment scene. In this embodiment, the kitchen scene data from the Mip-NeRF 360 dataset is used as the image data. The scene image sequence is preprocessed with size changes, color space conversions, etc., to obtain a set of input images of uniform specifications.

[0042] Step 2: Under a unified low-resolution configuration, input all views into the Fast3R multi-view feedforward network to obtain the local point map, global point cloud, and confidence information for each view. Since Fast3R's shared coordinate system is related to the initial view of the input sequence, an inappropriate selection of the initial view can affect subsequent global geometric stability. The specific implementation is as follows:

[0043] Step 2.1: Since local point maps typically have higher detail precision, a set of pixels with high confidence is selected from the local point map. For each view Confidence plot For sets The mean confidence score of the middle pixels is calculated to obtain the mean confidence score component:

[0044] + (1)

[0045] The higher the confidence component, the more reliable the effective region for geometric reasoning and the lower the tendency of point graph noise.

[0046] Step 2.2: To suppress local geometric noise that may occur in scenes with low texture, repetitive texture, or large flat surfaces, a depth scalar map is obtained on the aligned point map, and the relative gradient is calculated to finally obtain the stable depth component.

[0047] (2)

[0048] in, To align depth values, Representing local depth changes and using Perform scale normalization, and at the same time utilize The overall depth fluctuation is measured and penalized by the second term to obtain the depth stability component.

[0049] Step 2.3: Obtain the unit vector of the observation direction for each view from the coarse pose. and select its neighborhood set.

[0050] The average included angle is calculated and normalized to obtain the viewpoint diversity components:

[0051] (3)

[0052] in, and For the view and its neighborhood set The unit vector of the observation direction of the interior view is obtained through... The perspective diversity component is obtained by truncating to the [0,1] interval.

[0053] Step 2.4: After performing Min-Max normalization on the above three components, perform a weighted sum to obtain the view quality score:

[0054] (4)

[0055] in and , To prevent division by zero constants, This indicates that Min-Max normalization is performed on all views.

[0056] Step 3: Construct Anchor scores based on the low-resolution inference results, select the best Anchor view, and use it as the first frame to rearrange the sequence, generating an image list file (image_list) and a matching image pair file (Pairs) consistent with the rearrangement. Anchor optimization diagram is shown below. Figure 2 As shown, the specific implementation is as follows:

[0057] Step 3.1: To avoid selecting views that are "high quality but have little overlap with other views" as anchors, a mutual view coverage score is introduced for any pair of views. High-confidence 3D points are uniformly represented in a shared world coordinate system. ,Will Projecting the pixel coordinates onto view j yields the following:

[0058] (5)

[0059] The proportion of images falling within the valid image area is used as the view pair overlap. :

[0060] (6)

[0061] Based on this, the overlap between view i and other views is summed after threshold filtering. :

[0062] (7)

[0063] in This is the overlap threshold, used to determine valid overlap relationships.

[0064] Step 3.2: The optimal Anchor score is defined as the product of the quality score and the mutual coverage score, and an index is introduced. Adjusting co-visibility weights:

[0065] (8)

[0066] Select The largest view is used as the best Anchor view and placed at the beginning of the input sequence to rearrange the image sequence and generate an image_list file, thereby reducing the randomness caused by Fast3R's "first view defines the global coordinate system".

[0067] Step 3.3: For each view Based on mutual visual overlap Sort the views from largest to smallest and select the top-K views to form a candidate matching set. The process generates Pairs image pairs. Furthermore, to avoid isolated views in the candidate matching graph that would lead to a lack of constraints in subsequent optimization, graph connectivity is checked during the Pairs generation stage. When a view has no effective neighbors, it is forced to establish a pairing edge with the hub view with the highest degree, thereby ensuring the connectivity and optimization stability of the global constraint graph.

[0068] Step 4: Under high-resolution configuration, Fast3R inference is performed on the rearranged image sequence to obtain a more refined point map and confidence level. Subsequently, using the pixel and 3D correspondence provided by the point map, the extrinsic parameters of the camera for each view are solved within the iterative PnP framework. The camera focal length can be estimated by the network's prediction of the first view and used as the initial value for other views. After implementing SfM, its pose and intrinsic parameters are read as initial values ​​and fused with the Fast3R-estimated intrinsic parameters according to weights to improve stability.

[0069] Step 5: To improve the accuracy of camera intrinsic and extrinsic parameters and suppress noise, this invention, based on the initial camera parameters, primarily uses "dense feature consistency constraints between adjacent views" and secondarily uses "reprojection constraints of SfM sparse trajectories," thereby improving overall geometric consistency while maintaining computational controllability. A schematic diagram of the joint geometric optimization is shown below. Figure 3 As shown, the specific implementation is as follows:

[0070] Step 5.1: Construct sparse reprojection constraints. Based on the Pairs image pair constraints generated in Step 3.3, perform SIFT feature extraction and matching on the view pairs to obtain sparse corresponding points. Then, perform geometric verification on the matching points and triangulation to obtain sparse 3D points. Construct sparse reprojection error:

[0071] (9)

[0072] in For the first The three-dimensional point at the th t The observed pixels on the image For projection function, This is the Charbonnier robust kernel function.

[0073] Step 5.2: Continue using the Pairs file generated in Step 3.3 as the view adjacency graph, only on adjacent edges. Establish cross-view consistency constraints to reduce computational overhead and avoid the propagation of erroneous constraints caused by non-overlapping views.

[0074] Step 5.3: Construct dense consistency constraints in the source view. Above, pixels are sampled from high-confidence regions with significant gradients. An inverse depth parameter ρ is introduced to the sampling points to allow for fine-tuning of the depth along the line-of-sight direction, and a multi-scale feature pyramid is constructed on the source view and sampled. Then, feature metrics are compared to construct a dense consistency term:

[0075] (10)

[0076] Among them, weight It can be determined jointly by confidence and gradient:

[0077] (11)

[0078] Step 5.4: Constructing the joint optimization objective and solving:

[0079] (12)

[0080] By using gradient-based iterative optimization to update the camera pose and inverse depth parameters of sampling points on a multi-scale pyramid from coarse to fine, and fixing the pose increment of the Anchor view to zero to eliminate the canonical degrees of freedom, the convergence stability is improved, resulting in a more accurate and consistent set of camera parameters, providing high-quality input for 3DGS training.

[0081] Step 6: Summarize the optimized camera parameters and point maps from Step 5 to obtain a candidate point cloud set. Allocate a sampling budget to each view based on the overall view quality and prioritize high-quality points within the pixel grid. Finally, export the initialization data, including a sparse and uniform high-quality point cloud. The QSP strategy diagram is shown below. Figure 4 As shown. The specific implementation is as follows:

[0082] Step 6.1: Construct a pixel-level quality map for each view pixel. Construct quality score Taking into account confidence level, local-global consistency, self-weight projection error, cross-viewpoint depth consistency, and curvature information, it can be defined as a multiplicative coupling form:

[0083]

[0084] (13)

[0085] in This represents the distance between the world point obtained by pose transformation of the local point map and the global point map at this pixel. To account for the deviation between the point image and the original pixel coordinates after projecting it back onto the pixel grid, For cross-perspective consistency, As an approximate measure of curvature, As weight.

[0086] Step 6.2: By View Quality Perform sampling budget allocation and divide each image into a regular grid, selecting within each grid. The largest pixel corresponds to a 3D point as a candidate point. When the number of candidate points is insufficient, higher quality points are added from the entire image to ensure coverage and balance within the view.

[0087] Step 6.3: Map all candidate points to a 3D voxel mesh, sort them in descending order of quality score within each voxel, and retain only the Top-K points. The output point cloud exhibits an approximately uniform spatial distribution, which ensures both geometric structure and texture details while significantly reducing the size of the point cloud.

[0088] Step 7: Export in COLMAP format and use for 3DGS initialization. Export the camera-inside and outside point cloud obtained in Step 5, along with the initial point cloud obtained in Step 6, as COLMAP standard format data. This allows for direct integration into mainstream 3DGS rendering, improving convergence speed and final rendering quality.

[0089] The Mip-NeRF 360 dataset contains multiple indoor and outdoor scenes and is commonly used to evaluate novel perspective synthesis and 3D reconstruction tasks of neural radiation fields in unbounded scenes. The rendering accuracy results for the kitchen scene are shown in Table 1.

[0090] Table 1 Comparison of rendering result metrics for different models

[0091] Model PSNR↑ SSIM↑ LPIPS↓ CF-3DGS 26.63 0.794 0.373 ZeroGS 30.45 0.919 0.131 COLMAP+3DGS 30.18 0.917 0.134 Mip-Splatting 31.21 0.935 0.116 EDGS 31.14 0.926 0.121 Scaffold-GS 31.02 0.925 0.127 Fast3R+3DGS 22.64 0.572 0.623 Fast3R+Anchor+3DGS 25.85 0.721 0.382 Fast3R+Anchor+BA+3DGS 30.97 0.923 0.129 Fast3R (Full) + 3DGS 31.23 0.939 0.114

[0092] Comparison of ablation experiment rendering effects as shown in the image. Figure 5 As shown.

[0093] Based on the standardized camera parameters and optimized point cloud results described above, this invention is directly applied to the digital practice of 3D clothing stores in the embodiments. It efficiently completes high-precision 3D reconstruction and immersive display of clothing products and store displays. Its practical application effects can be referenced. Figure 6 As shown.

[0094] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A three-dimensional Gaussian sputtering initialization method based on Fast3R, characterized in that, The method includes the following steps: S1 acquires the scene image sequence for 3D Gaussian sputtering initialization, preprocesses the scene image sequence, and obtains a set of input images of uniform specifications; S2: Perform low-resolution Fast3R inference on the input image set to obtain the global point map, local point map and confidence map corresponding to each view, and construct view quality and mutual view correlation information based on the inference results; S3: Calculate the Anchor score of each view based on view quality and mutual view correlation information, select the best Anchor view from the input image set; rearrange the image sequence with the best Anchor view as the first frame, generate an image list file consistent with the rearrangement and generate a matching constraint file; S4: Perform high-resolution Fast3R inference on the rearranged image sequence, and use iterative PnP with RANSAC to solve the initial camera extrinsic parameters for each view; S5: Construct a joint geometric optimization model: Using the initial camera parameters and Fast3R point map as initialization priors, and combining feature extraction and matching, sparse reprojection constraints formed by triangulation, and dense consistency constraints formed by cross-view reprojection of point map, robust joint optimization of camera extrinsic parameters is performed to obtain the camera parameter set. S6: Construct candidate point clouds based on camera parameter sets and point map results, introduce QSP strategy, prioritize the retention of high-quality point clouds while satisfying the constraint of uniform spatial distribution, and output the initial point cloud for 3DGS training; S7: Convert and export the camera parameter set and the initial point cloud as input for the 3DGS task.

2. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 1, characterized in that, The low-resolution Fast3R inference described in S2 includes: All view images are input into the Fast3R network to obtain the local point map, global point map, confidence map, and coarse camera pose estimate for each view. The confidence map is statistically analyzed in the spatial dimension to obtain the view confidence mean, which is used as the first component of view quality. The mutual viewing intensity between views is estimated based on the point map visibility and overlapping area of ​​any two views, and a mutual viewing association table is formed accordingly.

3. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 2, characterized in that, The specific calculation steps for the Anchor score described in S3 are as follows: Anchor scores include a view quality item and a mutual visibility item: the view quality item is obtained by weighting the view confidence statistic, depth stability and view diversity, and is used to measure the geometric reliability of each view; the mutual visibility item is obtained by the percentage overlap between the view and other views, and the higher the percentage overlap, the greater the mutual visibility. Anchor scoring combines quality and mutual visibility factors with preset weights, selecting the view with the highest score as the best anchor view to reduce global coordinate system bias and pose instability caused by random selection in the first frame.

4. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 3, characterized in that, The overlap ratio statistics of the mutual visibility item include: Select a local point map of the view and map it to a shared coordinate system through a local-global alignment transformation. Project the mapped 3D points onto the image plane of other views. Calculate the proportion of the projection falling within the effective area of ​​the image as the view pair overlap ratio. Aggregate the overlap ratios of each view pair to obtain the mutual visibility term of the view.

5. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 4, characterized in that, The matching constraint file mentioned in S3 includes selecting the top-K adjacent views with the highest overlap ratio in the mutual view association table for each view to form a candidate matching pair, and writing the candidate matching pair into the Pairs constraint file to reduce invalid matching and reduce the amount of subsequent optimization calculations.

6. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 5, characterized in that, The specific steps of S4 include: The initial camera parameter set is solved by focal length candidate search and RANSAC-PnP: a focal length candidate set is set for each view, and the corresponding intrinsic parameter matrix is ​​calculated for each candidate focal length; under the intrinsic parameter matrix, the correspondence between high-confidence 3D points and their 2D pixels is selected as PnP input based on the confidence map; RANSAC-PnP is used to solve the camera extrinsic parameters, and the focal length candidate with the smallest reprojection error is used as the optimal focal length and extrinsic parameter output for that view; when the image sequence comes from the same physical camera, the optimal focal length of the Anchor view is used as the globally shared focal length.

7. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 6, characterized in that, The sparse reprojection constraint term construction described in S5 includes: Based on the Pairs images generated by S3, feature extraction and matching are performed on the files to obtain sparse corresponding points across views. The above sparse corresponding points are combined with the camera parameters estimated by Fast3R as priors to perform triangulation to obtain sparse 3D points, and a reprojection error objective function is constructed as a sparse constraint term for the joint optimization objective.

8. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 7, characterized in that, The construction of the dense consistency constraint term includes: Several high-quality pixels are sampled on the point map of the source view to obtain their three-dimensional coordinates; the three-dimensional coordinates are projected onto the target view using the current camera parameters to obtain the projection position; corresponding features are sampled on the multi-scale image features of the source view and the target view respectively, feature consistency error is constructed and accumulated in a robust manner to form a dense consistency constraint term; the sparse reprojection constraint term and the dense consistency constraint term are weighted and summed as a joint optimization objective to iteratively update the camera extrinsic parameters.

9. The three-dimensional Gaussian sputtering initialization method based on Fast3R according to claim 8, characterized in that, The QSP strategy described in S6 includes uniform sampling within the view and global voxel constraints: For each view, a pixel-level quality map is constructed based on the confidence map, reprojection residual, local-global alignment residual, and gradient curvature information; the image is divided into a regular grid, and the 3D point corresponding to the pixel with the highest quality score in each grid is selected as a candidate point, and high-quality candidate points are supplemented from the entire image when there are insufficient candidate points; all view candidate points are mapped to a 3D voxel grid, and only the top-K candidate points are retained in each voxel according to their quality scores to obtain a spatially uniform and controllable sampling point cloud.

Images