Three-dimensional reconstruction method, system and equipment based on compressed symbol distance field and medium
By using the method based on the compressed symbol distance field, the problem of insufficient reconstruction accuracy of 3D Gaussian splashing in infinite boundary scenes is solved, and high-precision and high-quality three-dimensional scene reconstruction is achieved, suitable for large-scale or boundary-free scenes.
Patent Information
- Application Number
- CN202411005855.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-22
AI Technical Summary
The existing 3D Gaussian splashing technology is difficult to achieve accurate geometric reconstruction in infinite boundary scenarios, resulting in poor accuracy in large-scale scenario reconstruction.
Using a three-dimensional reconstruction method based on compressed symbol distance field, a three-dimensional point cloud is generated by acquiring multi-view images and camera poses, a radiation field is constructed and differentiable rendering is performed, the parameters of two-dimensional Gaussian points are extracted, and the Gaussian bounding box optimization and dynamic programming are used. The algorithm optimizes the surface elements, and finally extracts the three-dimensional mesh model by compressing the symbol distance field for reconstruction.
The accuracy and visual quality of three-dimensional scene reconstruction are improved, and accurate reconstruction of boundless or large-scale scenes can be achieved, ensuring that the surface elements fit the physical properties and visual characteristics of the real image.
Smart Images

Figure CN119006741A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer graphics, and in particular to a three-dimensional reconstruction method, system, device and medium based on compressed signed distance field. Background Art
[0002] In the field of computer graphics and computer vision, achieving realistic novel view synthesis (NVS) and accurate geometric reconstruction has been a long-term key goal. As an alternative to implicit representation and feature grid-based representation, 3D Gaussian Splatting (3DGS) is particularly attractive in NVS because it can produce high-resolution realistic views in real time. 3DGS has developed rapidly and has been widely used in anti-aliasing rendering, material modeling, dynamic scene reconstruction, and animatable avatar creation. However, due to the conflict between the design of the 3D Gaussian voxel model on full angle radiation and the thin layer characteristics of the surface, it is insufficient in capturing complex geometric structures, making it difficult for 3DGS to achieve accurate geometric reconstruction in infinite boundary scenes, resulting in poor accuracy when reconstructing large-scale scenes. Therefore, it is necessary to provide a three-dimensional reconstruction method, system, device and medium based on compressed signed distance field. Summary of the invention
[0003] The present invention provides a three-dimensional reconstruction method based on a compressed signed distance field, which improves the problem of low accuracy in the prior art when large-scale scene reconstruction is achieved.
[0004] The present invention provides a three-dimensional reconstruction method based on a compressed signed distance field, comprising: obtaining different viewing angle images of a scene to be reconstructed and camera postures of corresponding viewing angles to generate a three-dimensional point cloud; performing radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes a plurality of two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud; extracting parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding face element; based on a Gaussian bounding box optimization algorithm and a deep warped back propagation algorithm of dynamic programming, performing performance optimization and depth adjustment on each face element to obtain an optimized face element; based on the compressed signed distance field, extracting a three-dimensional grid model from all the optimized face elements, and constructing a corresponding three-dimensional scene based on all the extracted three-dimensional grid models.
[0005] In one embodiment of the present invention, the radiation field modeling and differentiable rendering processing are performed on the three-dimensional point cloud to obtain the radiation field, including: performing radiation field modeling on the three-dimensional point cloud to generate an initial radiation field; wherein the initial radiation field includes multiple two-dimensional Gaussian points, each two-dimensional Gaussian point corresponds to a point cloud; initializing the parameters of each two-dimensional Gaussian point, and based on the differentiable rendering method, adjusting the parameters of each two-dimensional Gaussian point to obtain an optimized radiation field.
[0006] In one embodiment of the present invention, the steps of extracting the parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding face element include: extracting the parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding initial face element; performing perspective correction and rendering processing on each initial face element to obtain a final face element.
[0007] In one embodiment of the present invention, the perspective correction and rendering processing of each initial face element to obtain the final face element includes: performing perspective projection and correction processing on each initial face element to obtain a corrected face element; adjusting the coordinates of each corrected face element based on the ray-ellipse intersection method to obtain an adjusted face element; performing anti-aliasing processing on each adjusted face element to obtain an anti-aliased face element; and performing rasterization processing on each anti-aliased face element to obtain the final face element.
[0008] In one embodiment of the present invention, the Gaussian bounding box optimization algorithm and the depth-warped back-propagation algorithm based on dynamic programming perform performance optimization and depth adjustment on each face element to obtain an optimized face element, including: calculating the radius of the Gaussian bounding box of each face element based on the opacity of each face element; initializing the depth and weight of each face element, and calculating the depth loss of each face element; and optimizing the depth and weight of the corresponding face element according to each depth loss to obtain the corresponding optimized face element.
[0009] In one embodiment of the present invention, the radius of the Gaussian bounding box is Among them, α is the opacity and r is the radius of the Gaussian bounding box.
[0010] In one embodiment of the present invention, the method of extracting three-dimensional mesh models from all optimized face elements based on a compressed signed distance field, and constructing a corresponding three-dimensional scene based on all the extracted three-dimensional mesh models, includes: determining the truncated signed distance of the mesh points based on the depths of the preset mesh points in different optimized face elements, and generating a signed distance field; extracting isosurfaces from the signed distance field using a marching cubes algorithm to obtain three-dimensional mesh models corresponding to each optimized face element; merging all the three-dimensional mesh models into an overall three-dimensional mesh model, and constructing a corresponding three-dimensional scene based on the overall three-dimensional mesh model.
[0011] In one embodiment of the present invention, determining the truncated signed distance of the grid point according to the depth of the preset grid point in different optimized bins and generating the signed distance field includes: normalizing all the preset grid points; for each grid point: mapping the grid point to the world coordinate system through an inverse transformation; determining the truncated signed distance of the grid point according to the depth of the grid point in different optimized bins; and summarizing the truncated signed distance of each grid point to obtain the signed distance field.
[0012] In one embodiment of the present invention, a three-dimensional reconstruction system based on a compressed signed distance field is also provided, and the system includes: a data acquisition module, which is used to obtain different perspective images of the scene to be reconstructed and the camera postures of the corresponding perspectives to generate a three-dimensional point cloud; a radiation field construction module, which is used to perform radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes multiple two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud; a facet modeling module, which is used to extract the parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding facet; an optimization module, which is used to perform performance optimization and depth adjustment on each facet based on a Gaussian bounding box optimization algorithm and a dynamic programming deep warped back propagation algorithm to obtain an optimized facet; a reconstruction module, which is used to extract the three-dimensional mesh model from all the optimized facets based on the compressed signed distance field, and construct a corresponding three-dimensional scene based on all the extracted three-dimensional mesh models.
[0013] In one embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the electronic device implements any of the above-mentioned three-dimensional reconstruction methods based on compressed signed distance fields.
[0014] In one embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor of a computer, the computer executes any one of the above-mentioned three-dimensional reconstruction methods based on compressed signed distance fields.
[0015] The present invention proposes a three-dimensional reconstruction method based on a compressed signed distance field, which generates a three-dimensional point cloud by acquiring multi-view images of the scene to be reconstructed and the corresponding camera poses, and constructs a radiation field through these point clouds. Each two-dimensional Gaussian point in the radiation field can characterize the position and attribute information of the corresponding point cloud, thereby obtaining the corresponding face element. These face elements are optimized for performance and depth adjustment through Gaussian bounding box optimization and dynamic programming deep warping back propagation algorithm, which can not only realize the reconstruction of borderless scenes, but also ensure that the face elements can better fit the physical properties and visual features of the real image. Then, through the compressed signed distance field method, the reconstruction of the three-dimensional scene is realized according to these optimized face elements. In this way, the accuracy and visual quality of the three-dimensional scene reconstruction are improved, and the three-dimensional reconstruction of large-scale or borderless scenes can be realized. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A schematic flow chart of a three-dimensional reconstruction method based on a compressed signed distance field provided by an embodiment of the present invention;
[0017] Figure 2 The display is a schematic diagram of the coordinates of a local two-dimensional Gaussian point and a screen in one embodiment of the present invention;
[0018] Figure 3 Shown is a schematic diagram of an overall scheme in one embodiment of the present invention;
[0019] Figure 4a The color diagram is shown as a two-dimensional elliptical disk with surface elements;
[0020] Figure 4b A schematic diagram showing the normals of a two-dimensional elliptical disk as a surface element
[0021] Figure 4c Shown is a schematic diagram of the colors of surface elements after two-dimensional Gaussian splashing in the present invention;
[0022] Figure 4d Shown is a schematic diagram of the normal lines of surface elements after two-dimensional Gaussian sputtering in the present invention;
[0023] Figure 5 It is a schematic diagram showing the triangulated meshing of the radiation field formed by the surface elements in one embodiment of the present invention;
[0024] Figure 6 Shown is a comparison of mesh extraction using different reconstruction algorithms;
[0025] Figure 7 The display is a comparison chart of reconstruction effects of different reconstruction algorithms;
[0026] Figure 8 Shown are mesh extraction result diagrams of different objects in one embodiment of the present invention;
[0027] Fig. 9 Shown are radiation field rendering results of different objects in one embodiment of the present invention;
[0028] Fig.10a Renderings shown are reconstructions of the bounded scene;
[0029] Fig.10b The triangle mesh image reconstructed for the scene with boundaries is shown;
[0030] Fig.10c Renderings shown reconstructed for a boundless scene;
[0031] Fig.10d Displays the reconstructed triangular mesh for an unbounded scene;
[0032] Fig.11 Shown is a structural block diagram of a three-dimensional reconstruction system based on a compressed signed distance field provided by an embodiment of the present invention;
[0033] Fig.12 Shown is a structural schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0034] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0035] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0036] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.
[0037] The inventors found that although the three-dimensional Gaussian splatting (3DGS) technology has made a breakthrough in radiation field reconstruction, it can achieve high-quality new perspective synthesis and fast rendering speed. However, due to the multi-perspective inconsistency of three-dimensional Gaussian, 3DGS has limitations in accurately representing surfaces. In addition, surface elements (surfels, i.e., surface elements) are an effective method for representing complex geometric structures. Surfels use shape and shadow properties to locally approximate the surface of an object and can be derived from known geometric structures. These surface elements are widely used as effective geometric representation methods in SLAM and other robotic tasks. Subsequent progress has incorporated surface elements into a differentiable framework. However, these methods usually require real geometric information of the ground, depth sensor data, or operate in restricted scenes under known lighting conditions. In view of the above limitations, the present invention proposes a differentiable surfel splatting method, which is a new geometrically accurate radiation field modeling and reconstruction method. The present invention overcomes the limitations of previous technologies, especially the shortcomings in accurately capturing detailed geometric structures, while maintaining real-time rendering and high-quality visual effects.
[0038] The present invention provides a 3D reconstruction method based on compressed signed distance field, which is a new strategy for reconstructing geometrically accurate radiation fields from multi-view images. Unlike 3D Gaussians, 2D Gaussians essentially provide perspective-consistent geometry for surface modeling. Therefore, the core idea of the present invention is to compress the 3D volume into a set of 2D oriented planar Gaussian disks. In order to accurately restore thin surfaces and achieve stable optimization, the present invention introduces a perspective-accurate facet splatting process that utilizes ray-facet intersection and rasterization. In addition, to further improve the reconstruction quality, depth distortion and normal consistency conditions are added. The differentiable rendering method of the present invention can perform noise-free and detailed geometric reconstruction while maintaining competitive appearance quality, fast training speed and real-time rendering.
[0039] See also Figure 1 , the 3D reconstruction method based on compressed signed distance field includes the following steps:
[0040] S1. Obtain images of the scene to be reconstructed from different perspectives and the camera poses of the corresponding perspectives to generate a three-dimensional point cloud.
[0041] Images of the scene are captured from multiple angles, and the camera pose information at each viewpoint is obtained through camera parameters. The camera pose refers to the specific position and orientation of the camera in space. Feature points are extracted from each image and matched between images at different viewpoints. A disparity map is generated based on the matched feature points and camera pose. The disparity map is converted into three-dimensional coordinates using triangulation methods to form a three-dimensional point cloud.
[0042] S2. Perform radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes multiple two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud.
[0043] In one embodiment of the present invention, radiation field modeling and differentiable rendering processing are performed on the three-dimensional point cloud to obtain the radiation field, including:
[0044] Perform radiation field modeling on the three-dimensional point cloud to generate an initial radiation field; wherein the initial radiation field includes a plurality of two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud;
[0045] Initialize the parameters of each two-dimensional Gaussian point, and adjust the parameters of each two-dimensional Gaussian point based on the differentiable rendering method to obtain the optimized radiation field.
[0046] Each point in the three-dimensional point cloud is converted into a two-dimensional Gaussian point, and these two-dimensional Gaussian points together form the initial radiation field. The corresponding parameters are initialized for each two-dimensional Gaussian point, and the parameters of the two-dimensional Gaussian point include the position, scale, direction, opacity and color of the two-dimensional Gaussian point. Based on the differentiable rendering method, each two-dimensional Gaussian point in the initial radiation field is rendered to generate a rendered image. The loss between the rendered image and the real image of the scene to be reconstructed taken from different perspectives is calculated. According to the calculated loss, the parameters of the two-dimensional Gaussian points are updated through the back propagation algorithm. These updated two-dimensional Gaussian points constitute the optimized radiation field.
[0047] S3. Extract the parameters of each two-dimensional Gaussian point from the radiation field and construct the corresponding surface element.
[0048] In one embodiment of the present invention, the parameters of each two-dimensional Gaussian point are extracted from the radiation field to construct the corresponding facet, including two steps S31 and S32, which specifically include:
[0049] S31. Extract the parameters of each two-dimensional Gaussian point from the radiation field and construct the corresponding initial surface element.
[0050] Extract the parameters of each two-dimensional Gaussian point from the radiation field. The extracted parameters of the two-dimensional Gaussian point include the center point p k , two main tangent vectors t u and t v , and the scaling vector S = (s u ,s v ), these parameters together define the position and shape of each two-dimensional Gaussian point in three-dimensional space. The center point p k Represents the position of the two-dimensional Gaussian point in three-dimensional space, and the tangent vector t u and t vRepresents the direction of the two-dimensional Gaussian point in the local coordinate system. The scaling vector S is used to adjust the size of the two-dimensional Gaussian point to affect its diffusion in the local coordinate system. u and t v Multiply them together to get the normal vector t w =t u ×t v , by the normal vector t w and two tangent vectors t u and t v The 3×3 rotation matrix R=[t u , t v , t w ]. In addition, the scaling factor s u and v Arranged into a 3×3 diagonal matrix S with the last entry being zero. It can be understood that the center point p here k , scaling factor s u and v , tangent vector t u and t v is an initialized value, which can be updated iteratively according to the value of the loss function until the face element reaches the optimal state.
[0051] In order to transform the above two-dimensional Gaussian point from the local tangent plane to the three-dimensional world coordinate system, the rotation matrix R and the center point p k Construct a transformation matrix H, as shown in formula (1):
[0052]
[0053] H is the transformation matrix, which is a 4× 4 homogeneous transformation matrix used to describe the geometric shape of two-dimensional Gaussian points. u and v is the scaling factor, t u and t v is the tangent vector of the two-dimensional Gaussian point in the local coordinate system, which is used to describe the specific position of the two-dimensional Gaussian point in the local coordinate system. k is the center point of the two-dimensional Gaussian point, R is the rotation matrix, and S is the scaling vector. u t u 、s v t v Represents vector t u and t v Multiply by the corresponding scaling factor s u and v , thereby realizing the rotation and scaling of the local coordinate system to ensure the accurate transformation of the coordinates in the two coordinate systems. Through the transformation matrix, the two-dimensional Gaussian point is transformed from the local coordinates to the world coordinates as shown in formula (2):
[0054] P(u,v)=p k +s u t u u+s v t v v=H(u,v,1,1) T (2)
[0055] Through formula (2), the two-dimensional Gaussian point can be transformed from the local coordinate system to the global coordinate system. P(u, v) is the position of the local coordinates u and v in the world coordinate system. This position reflects the geometric and visual properties of the original two-dimensional Gaussian point in the three-dimensional space. Therefore, through this transformation, each two-dimensional Gaussian point is not only a simple point, but is converted into a three-dimensional surface element with a certain shape, direction and scale. This surface element is the initial surface element corresponding to the two-dimensional Gaussian point.
[0056] S32, performing perspective correction and rendering processing on each initial face element to obtain a final face element.
[0057] For a point u=(u, v) in the local coordinate system, the two-dimensional Gaussian value G(u) of the point can be evaluated by the standard Gaussian, as shown in formula (3):
[0058]
[0059] In addition, following the 3DGS, each two-dimensional Gaussian ellipse not only has opacity α, but also uses spherical harmonics to parameterize its view-dependent appearance parameter c. The spherical harmonics dynamically adjust the appearance to ensure that the visual effect of the ellipse can change according to the change of the viewing angle, so as to ensure the accuracy of the subsequent three-dimensional scene reconstruction. In addition, the accuracy of the surface element is improved by rendering the surface element with the correct perspective, and the surface element after the above processing is the final surface element.
[0060] In one embodiment of the present invention, perspective correction and rendering are performed on each initial face element to obtain a final face element, including four steps from S321 to S324, specifically:
[0061] S321, performing perspective projection and correction processing on each initial face element to obtain a corrected face element.
[0062] When rendering a face, it is first necessary to perform an affine approximation of the perspective projection on each two-dimensional Gaussian primitive, so as to map the Gaussian primitive into the image space. However, this method is more accurate at the center of the Gaussian, but the approximation error increases as it moves away from the center. To improve the accuracy, the present invention adopts a formula based on homogeneous coordinates to convert the coordinates of the world space into the screen space through the W matrix to obtain the position of the midpoint in the screen space, as shown in formula (4):
[0063] x=(xz,yZ,z,1) T =WP(u,v)=WH(u,v,1,1) T (4)
[0064] W is the transformation matrix from world space to screen space, and x is the mapping position of the two-dimensional Gaussian point from three-dimensional space to screen space, which represents a homogeneous ray emitted from the camera, passing through the pixel (x, y) and intersecting the surface element at depth z. In this way, the corrected surface element can be obtained.
[0065] S322. Adjust the coordinates of each corrected face element based on the ray-ellipse intersection method to obtain an adjusted face element.
[0066] Considering the existing rasterization process, M = (WH) - 1 The implicit method of projecting the quadratic surface of the Gaussian primitive into the screen space, but the inventors found that when the ellipse degenerates into a line segment (i.e., when viewed from the side), this inverse transformation will lead to data instability. To improve this problem, the present invention adopts an explicit ray-ellipse intersection method, which calculates the rays intersecting the ellipse to determine the intersection point, thereby avoiding the use of unstable inverse transformation.
[0067] Specifically, when using the ray-ellipse intersection method, in order to efficiently determine the intersection of the ray and the ellipse, first find the intersection of three non-parallel planes. Given an image coordinate x = (x, y), the ray passing through the pixel is parameterized as the intersection of two orthogonal planes: the x plane and the y plane. Specifically, the x plane is defined by the normal vector (-1, 0, 0) and the offset x, so it can be represented as a four-dimensional homogeneous plane hx = (-1, 0, 0, x) T Similarly, the y plane can be defined as hy = (0, -1, 0, y) T . Therefore, we know that the ray x = (x, y) is determined by the intersection of the x plane and the y plane. Then, transform these two planes to the local coordinate system of the two-dimensional Gaussian primitive (i.e., the uv coordinate system). Since the transformation matrix M is used to transform the points on the plane, it is equivalent to using its inverse transposition M -T To transform the homogeneous plane parameters, therefore, M = (WH) -1 Equivalent to (WH) T , in this way, the explicit matrix inversion process is avoided. Through the above changes, we get formulas (5) and (6):
[0068] h u =(WH) T h x (5)
[0069] h v =(WH)T h y (6)
[0070] Through the above formula, the two-dimensional Gaussian point on the two-dimensional Gaussian plane is expressed in the form of (u, v, 1, 1), and the intersection point falls on the transformed x-plane and y-plane, so the condition is satisfied, as shown in formula (7):
[0071] h u (u, v, 1, 1) T =h v (u, v, 1, 1) T =0 (7)
[0072] According to formula (7), u(x) and v(x) can be effectively solved to obtain u(x) (i.e., the position of the screen space mapped to the local coordinate system), as shown in formulas (8) and (9):
[0073]
[0074] are the i-th parameters of the four-dimensional homogeneous plane parameters respectively. The depth z of the intersection point is obtained according to formula (4), and the Gaussian value is evaluated using formula (3), thereby completing the whole process and obtaining the adjusted surface element.
[0075] like Figure 2 As shown, for a two-dimensional Gaussian point (u, v) in the local coordinate system, it presents an ellipse, each ellipse has a center point p k Controls the size and shape of the ellipse depending on the scaling factor s u and v , and the vector t u and t v When the Gaussian points in the local coordinate system are transformed to the screen coordinate system, the Gaussian points are mapped to the preset grid points.
[0076] S323, performing anti-aliasing processing on each adjusted face element to obtain an anti-aliased face element.
[0077] Considering three-dimensional rendering, when observing two-dimensional Gaussian distributed surface elements from an oblique perspective, these surface elements may degenerate into a line in the screen space, which may be omitted or cannot be correctly represented during the rasterization process. To improve this problem, this application uses anti-aliasing to improve this visual effect. Anti-aliasing is to reduce or eliminate the visual distortion caused by insufficient sampling. For two-dimensional Gaussian surface elements, it is to improve the above situation by smoothing their representation in the screen space. Specifically, this application uses an object space low-pass filter as shown in formula (10):
[0078]
[0079] By using formula (8) and (9), we can get u(x), where c is the center point p k The projection of is the lower bound of a fixed screen-space Gaussian low-pass filter with center c and radius σ. σ is This ensures that there are enough pixels to cover during the rendering process, thereby improving the visual distortion problem and obtaining anti-aliased surfaces.
[0080] S324, performing rasterization processing on each anti-aliased surface element to obtain a final surface element.
[0081] Rasterization is the process of converting two-dimensional Gaussian surfaces into pixel graphics that can be displayed on the screen. It is used to convert three-dimensional scene data into two-dimensional images. Before rasterization, all two-dimensional Gaussian surfaces are sorted according to the depth of their centers to ensure that objects farther away from the observer do not cover nearby objects. The screen is divided into a preset number of regions, each of which is called a tile, and each region processes the surface elements independently.
[0082] In addition, this application also uses a volume blending method to integrate the appearance of each facet, as shown in formula (11):
[0083]
[0084] c(x) is the color value at the final calculated screen coordinate x, c i is the color value of the i-th face element, α i , α j are the opacities of the i-th and j-th face elements, respectively. are the Gaussian values of the i-th and j-th face elements at the screen coordinate x after anti-aliasing processing. During rasterization, face elements are rendered from front to back. When the cumulative opacity of a certain area reaches saturation, the iteration of the area is terminated and the rendering process of the area is stopped. When all areas are rendered, all face elements are processed and the final face element is obtained.
[0085] Furthermore, in order to achieve a more realistic 3D reconstruction effect, the present invention adjusts the facet parameters so that the facets can better match the visual information of the real world. Specifically, in order to minimize the distance between the ray and the spot intersection to optimize the weight distribution, the present application proposes a depth distortion loss L d , as shown in formula (12):
[0086] L d =∑ i,j w i w j |z i -z j | (12)
[0087] is the mixing weight of the ith intersection, z i 、z j are the depths of the i-th and j-th intersection points, respectively. is the mixing weight of the j-th intersection.
[0088] Considering that this application is based on two-dimensional Gaussian surface elements, it is necessary to ensure that all Gaussian spots are locally aligned with the actual surface. In volume rendering, multiple semi-transparent surface elements may exist along the ray. Consider the end point intersection p when the cumulative opacity reaches 0.5. s The actual surface at the location is calculated by comparing the normal of the spot with the normal consistency loss L of the surface normal calculated by the depth map gradient. n , to correct these bins, as shown in formula (13):
[0089]
[0090] i is the index value of the intersection point along the ray, w i is the mixing weight of the i-th intersection, n i is the normal of the light spot pointing to the camera, and N is the surface normal estimated based on the depth map gradient. Specifically, the calculation method of the normal N is shown in formula (14):
[0091]
[0092] N(x,y) is the unit surface normal vector at coordinates (x,y), are the gradient vectors in the x and y directions respectively. By calculating the normal consistency loss, the normal vector of the face element is adjusted so that the normal of the face element is as close to the normal of the real geometric surface as possible, thus presenting a more realistic visual effect.
[0093] Finally, a set of images with known poses are used to optimize the model generated by the initial sparse point cloud, and the total loss function L used is shown in formula (15):
[0094] L=L c +αL d +βL n (15)
[0095] L c It is the RGB reconstruction loss, which combines the L1 norm and the structural similarity metric to measure the similarity between the reconstructed image and the real image. α and β are adjustment parameters used to balance the influence of different loss terms. By continuously adjusting the facet parameters in this way, when the loss function is minimized, the facet is considered to have reached the optimal state.
[0096] S4. Based on the Gaussian bounding box optimization algorithm and the deep warped back propagation algorithm of dynamic programming, the performance of each face element is optimized and the depth is adjusted to obtain the optimized face element.
[0097] In one embodiment of the present invention, based on the Gaussian bounding box optimization algorithm and the deep warping back propagation algorithm of dynamic programming, the performance of each face element is optimized and the depth is adjusted to obtain the optimized face element, including three steps S41, S42, and S43, which specifically include:
[0098] S41. Calculate the radius of the Gaussian bounding box of each face element based on the opacity of each face element.
[0099] Since real-life scenes are usually more complex, expressing a realistic scene usually requires a two-dimensional Gaussian with millions of points. At the same time, in order to be able to render high-resolution images in real time, alignment is required to perform necessary algorithm pruning and hardware acceleration. One way to optimize the rendering process is to use a Gaussian bounding box to reduce unnecessary calculations. The Gaussian bounding box is an area surrounding a two-dimensional Gaussian distribution, which ignores the peripheral area with extremely small contributions. When calculating the bounding box of a two-dimensional Gaussian, traditional methods usually set a fixed threshold (such as r = 3 standard deviations) and only consider information within the range of three standard deviations extending from the center to the outside of the Gaussian function. However, this method does not take into account the transparency of the Gaussian elements, resulting in overly conservative calculations. To improve this situation, the present application uses a Gaussian bounding box radius calculation method that takes transparency into account. According to αexp(-r 2 / 2)=exp(-3 2 / 2), we can see that the radius of the bounding box is calculated as shown in formula (16):
[0100]
[0101] α is the opacity, and r is the radius of the Gaussian bounding box. Therefore, as the transparency decreases, the radius of the corresponding bounding box also decreases. By dynamically calculating the radius of the Gaussian bounding box, a more compact bounding box is provided for each Gaussian element, reducing unnecessary rendering and improving rendering speed.
[0102] S42: Initialize the depth and weight of each face element, and calculate the depth loss of each face element.
[0103] For each face element: its depth and weight are first initialized, and the overall depth error is calculated using the formula for depth distortion loss, as shown in formula (17):
[0104]
[0105] w i 、w jare the weights of the i-th and j-th face elements, respectively, and d i d j are the depths of the i-th and j-th facets respectively, and N is the total number of facets.
[0106] S43. According to each depth loss, the depth and weight of the corresponding surface element are optimized to obtain the corresponding optimized surface element.
[0107] Since each facet must be calculated according to formula (17), the complexity of the algorithm is O(n 2 ), n is the total number of facets. For a large number of facets (such as millions), this approach will be very time-consuming and inefficient. To improve this problem, the present application decomposes the depth loss into a term L directly associated with each Gaussian point k , as shown in formula (19):
[0108]
[0109] The weight and depth of Gaussian are derived with O(1) computational overhead, as shown in formulas (20) and (21):
[0110]
[0111] D=∑ i≠k w i d i , represents the weighted sum of the depth and weight of other Gaussian points except the current Gaussian point k, A = ∑ i≠k w i , represents the sum of the weights of other Gaussian points except the current Gaussian point k. The weight and depth of each Gaussian point are calculated by formulas (20) and (21), so that the optimized face element can be obtained through the weights and depths of these calculated Gaussian points.
[0112] S5. Based on the compressed signed distance field, extract the three-dimensional mesh models from all the optimized face elements, and construct the corresponding three-dimensional scene based on all the extracted three-dimensional mesh models.
[0113] In one embodiment of the present invention, based on the compressed signed distance field, a three-dimensional mesh model is extracted from all optimized facets, and a corresponding three-dimensional scene is constructed according to all the extracted three-dimensional mesh models, including:
[0114] S51 . Determine the truncated signed distance of the grid point according to the depth of the preset grid point in different optimized bins, and generate a signed distance field.
[0115] Before reconstructing a 3D scene, the grid structure of the scene is first pre-constructed in the 3D space. The grid structure consists of multiple evenly distributed grid points, which are used to obtain structural information in the 3D space. The position of each preset grid point is mapped to the optimized surface element model, and the truncated signed distance of each point is determined based on the depth information of these grid points in different surfaces. These truncated signed distances are aggregated together to form a signed distance field of the entire scene.
[0116] In one embodiment of the present invention, determining the truncated signed distance of the grid point according to the depth of the preset grid point in different optimized bins and generating the signed distance field includes:
[0117] Normalize all preset grid points;
[0118] For each grid point:
[0119] Map the grid points to the world coordinate system through inverse transformation;
[0120] Determine the cutoff signed distance of the grid point according to the depth of the grid point in the different optimized bins;
[0121] The truncated signed distance of each grid point is summed up to obtain the signed distance field.
[0122] When performing three-dimensional reconstruction, the radiation field must first be converted into a continuous implicit field. Currently, the conversion to the implicit field is usually done by TSDF fusion (truncated signed distance function fusion), Boisson reconstruction and other methods. However, the inventors have found that these methods can only work within a limited bounding box and cannot extract triangular meshes of large-scale scenes, such as modeling the background. In order to achieve the reconstruction of borderless or ultra-large scenes, the present invention extracts each grid point in the preset Euclidean space based on a compressed ball, and compresses the grid point into a sphere with a radius of 2 according to formula (22):
[0123]
[0124] In this way, all grid points are compressed into a standardized sphere, which can limit the data boundary that may have extended infinitely to a certain range, so that the reconstruction of a boundless or infinite scene can be realized. After the above normalization process, the compressed grid points are restored to the world coordinate system using an inverse transformation. The inverse transformation method is shown in formula (23):
[0125]
[0126] After the inverse transformation is completed, for each grid point x after the inverse transformation, it is converted into the actual space coordinate system. When the center coordinate c of the sphere and the radius r of the sphere are known, the real space coordinate x'=inv-contract(x)×r+c is obtained. The cutoff distance t' of each real space coordinate x' is also adjusted to the actual size according to the same inverse transformation rule, t'=inv-contract(t)×r, t is the predefined cutoff distance, and the position of each grid point x∈[-2,2] 3 Then, the truncated signed distance of each point is determined according to the depth information of the grid point in different optimized facet models. Specifically, for each grid point’s real spatial coordinate x’, query the grid point at each viewing angle D i The depth d under the camera is calculated according to the camera pose P i Calculate the point in view P i The predicted depth z under tsdf(x') = clamp((dz) / t', -1.0, 1.0), the truncated signed distance of the point is calculated. The TSDF values calculated at each view are summarized and the results of all views are averaged to obtain the final signed distance field tsdf(X).
[0127] S52. Use the marching cubes algorithm to extract isosurfaces from the signed distance field to obtain a three-dimensional mesh model corresponding to each optimized face element.
[0128] The Marching Cubes algorithm is used to identify and extract zero isosurfaces in 3D volume data. This zero isosurface is the surface of an object. The algorithm traverses the voxel data of the entire signed distance field and calculates the exact position where the isosurface crosses the voxel boundary through linear interpolation, thereby obtaining a 3D mesh model corresponding to each face element.
[0129] S53, merging all three-dimensional grid models into an overall three-dimensional grid model, and constructing a corresponding three-dimensional scene according to the overall three-dimensional grid model.
[0130] All acquired 3D mesh models are merged into an overall 3D mesh model. The smoothness of the connection between the 3D mesh models can be improved by eliminating duplicate vertices and smoothing, so as to obtain an overall 3D mesh model. By adding textures, lighting, etc. to the overall 3D mesh model, the corresponding 3D scene is obtained.
[0131] See also Figure 3The present invention first preprocesses the three-dimensional point cloud of the scene, such as reducing noise, format conversion, etc. The processed point cloud is splashed with the differentiable surface element proposed by the present invention to generate a radiation field composed of several two-dimensional Gaussian points. Through adaptive point cloud density control, image loss and regularization loss, the obtained surface element is more in line with the real scene, thereby improving the accuracy of three-dimensional scene reconstruction.
[0132] Figure 4a and Figure 4b The radiation field image and normal image obtained when the two-dimensional elliptical disk is used as the surface element are presented respectively. It can be seen from the figure that both the reconstructed radiation field and the reconstructed surface normal have uneven edges in the reconstructed images. Figure 4c and Figure 4d When the two-dimensional Gaussian sputtering proposed in the present invention is used for facet reconstruction, the radiation field image and normal image are obtained. As can be seen from the figure, whether it is the reconstructed radiation field or the reconstructed surface normal, the reconstructed image has smooth edges and is more in line with the real scene.
[0133] Figure 5 The figure shows an overall three-dimensional mesh model generated by three-dimensional reconstruction technology using the method of the present invention. As can be seen from the figure, the surface of the object is smooth and the edge structure is clear, which shows that the method of the present invention can effectively restore complex three-dimensional geometric shapes and generate high-quality models by retaining a large number of detailed features.
[0134] Figure 6 The comparison of reconstruction results of the same input data using different 3D reconstruction techniques is shown. As can be seen from the figure, when the input data (Input) is a house of Scan 24 or a teddy bear of Scan 105, using the 3DGS method, some structures and detail features are lost after reconstruction, and the surface is relatively rough. Using the SuGaR method, more structures and detail features are lost after reconstruction. Using the method of the present invention (Ours), the structure and detail features are retained after reconstruction, and the surface of the reconstructed object is smoother.
[0135] from Figure 7It can be seen that the last three columns in the figure are the 3D mesh models of each method. It can be seen that compared with the existing 3D reconstruction methods such as 3DGS and SuGaR, the 3D mesh model obtained by this method (Ours) of the present invention has more details and smooth edge transitions. The third column in the figure is the surface normal map (Ours(normal)) of the scene reconstructed by the present invention, and the second column in the figure is the color reconstruction map (Ours(color)) of the present invention. It can be seen that this method of the present invention shows higher accuracy and detail retention in terms of color and normal, making the reconstructed scene more consistent with the corresponding real image (Ground truth).
[0136] Figure 8 The mesh extraction results of different objects by the present invention are shown. It can be seen from the figure that no matter whether the external contour of the object is simple or not, the present invention can accurately capture the detailed features of the external surface of the object, thereby achieving higher quality three-dimensional reconstruction. Fig. 9 The images rendered by the present invention on the radiation field of different scenes are shown. It can be seen from the figure that the images rendered by the method of the present invention are more realistic. This method can be applied to various scenes, from daily necessities to the entire environment, and realistic three-dimensional reconstruction can be performed.
[0137] Fig.10a A rendering showing the reconstruction of the scene with boundaries, Fig.10b The triangular mesh image of the reconstruction of a scene with boundaries is shown, from which it can be seen that the existing methods can only achieve reconstruction of scenes with clear boundaries. For large-scale scenes or scenes without boundaries, the existing reconstruction methods cannot be used to reconstruct them. Fig.10c , Fig.10d The rendering and triangular mesh of the reconstruction of the borderless scene are shown respectively. It can be seen that the reconstruction of the scene without borders or in a large range can be achieved by using the method of the present invention.
[0138] See also Fig.11The three-dimensional reconstruction system 100 based on the compressed signed distance field includes: a data acquisition module 110, a radiation field construction module 120, a facet modeling module 130, an optimization module 140 and a reconstruction module 150. The data acquisition module 110 is used to obtain different perspective images of the scene to be reconstructed and the camera posture of the corresponding perspective to generate a three-dimensional point cloud. The radiation field construction module 120 is used to perform radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes a plurality of two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud. The facet modeling module 130 is used to extract the parameters of each two-dimensional Gaussian point from the radiation field to construct the corresponding facet. The optimization module 140 is used to optimize the performance and adjust the depth of each facet based on the Gaussian bounding box optimization algorithm and the deep warping back propagation algorithm of dynamic programming to obtain the optimized facet. The reconstruction module 150 is used to extract the three-dimensional mesh model from all the optimized facets based on the compressed signed distance field, and construct the corresponding three-dimensional scene according to all the extracted three-dimensional mesh models.
[0139] For the specific definition of the three-dimensional reconstruction system based on the compressed signed distance field, please refer to the definition of the three-dimensional reconstruction method based on the compressed signed distance field in the above text, which will not be repeated here. Each module in the above-mentioned three-dimensional reconstruction system based on the compressed signed distance field can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware format, or can be stored in the memory in the computer device in software format, so that the processor can call the operations corresponding to the above modules.
[0140] It should be noted that, in order to highlight the innovative part of the present invention, the present embodiment does not introduce modules that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other modules in the present embodiment.
[0141] See also Fig.12 The electronic device 1 may include a memory 12, a processor 13 and a bus, and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a three-dimensional reconstruction program based on a compressed signed distance field.
[0142] Among them, the memory 12 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (for example: SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 12 can be an internal storage unit of the electronic device 1, such as a mobile hard disk of the electronic device 1. In other embodiments, the memory 12 can also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1. Further, the memory 12 can also include both an internal storage unit of the electronic device 1 and an external storage device. The memory 12 can not only be used to store application software and various types of data installed in the electronic device 1, such as codes for three-dimensional reconstruction based on compressed signed distance fields, etc., but can also be used to temporarily store data that has been output or is to be output.
[0143] In some embodiments, the processor 13 may be composed of an integrated circuit, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and combinations of various control chips. The processor 13 is the control core (Control Unit) of the electronic device 1, and uses various interfaces and lines to connect various components of the entire electronic device 1, and executes or executes programs or modules stored in the memory 12 (such as a three-dimensional reconstruction program based on a compressed signed distance field, etc.), and calls data stored in the memory 12 to execute various functions of the electronic device 1 and process data.
[0144] The processor 13 executes the operating system and various installed application programs of the electronic device 1. The processor 13 executes the application programs to implement the steps in the above-mentioned three-dimensional reconstruction method based on compressed signed distance field.
[0145] Exemplarily, the computer program may be divided into one or more modules, one or more modules are stored in the memory 12, and executed by the processor 13 to complete the present application. One or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device 1. For example, the computer program may be divided into a data acquisition module 110, a radiation field construction module 120, a surface element modeling module 130, an optimization module 140, and a reconstruction module 150.
[0146] The above-mentioned integrated unit implemented in the form of a software function module can be stored in a computer-readable storage medium, and the computer-readable storage medium can be non-volatile or volatile. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a computer device, or a network device, etc.) or a processor to perform part of the functions of the three-dimensional reconstruction method based on the compressed signed distance field in each embodiment of the present application.
[0147] In summary, the present invention discloses a three-dimensional reconstruction method, system, device and medium based on compressed signed distance field. By acquiring multi-view images of the scene to be reconstructed and the corresponding camera poses, a three-dimensional point cloud is generated, and a radiation field is constructed through these point clouds. Each two-dimensional Gaussian point in the radiation field can characterize the position and attribute information of the corresponding point cloud, thereby obtaining the corresponding face element. These face elements are optimized and adjusted in depth through Gaussian bounding box optimization and dynamic programming deep distortion back propagation algorithm, which can not only realize the reconstruction of borderless scenes, but also ensure that the face elements can better fit the physical properties and visual features of the real image. Then, through the compressed signed distance field method, the reconstruction of the three-dimensional scene is realized according to these optimized face elements. In this way, the accuracy and visual quality of the three-dimensional scene reconstruction are improved, and not only the reconstruction of the scene with boundaries can be realized, but also the three-dimensional reconstruction of large-scale or borderless scenes can be realized. Therefore, the present invention effectively overcomes the various shortcomings in the prior art and has a high industrial utilization value.
[0148] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.
Claims
1. A three-dimensional reconstruction method based on compressed signed distance field, characterized in that: The method comprises: Obtain images of the scene to be reconstructed from different perspectives and the corresponding camera poses to generate a three-dimensional point cloud; Performing radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes a plurality of two-dimensional Gaussian points, and each two-dimensional Gaussian point corresponds to a point cloud; Extracting parameters of each two-dimensional Gaussian point from the radiation field to construct corresponding surface elements; Based on the Gaussian bounding box optimization algorithm and the deep warped back propagation algorithm of dynamic programming, the performance of each face element is optimized and the depth is adjusted to obtain the optimized face element; Based on the compressed signed distance field, the three-dimensional mesh models are extracted from all the optimized face elements, and the corresponding three-dimensional scene is constructed according to all the extracted three-dimensional mesh models.
2. The three-dimensional reconstruction method based on compressed signed distance field according to claim 1, characterized in that: The step of performing radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain the radiation field includes: Performing radiation field modeling on the three-dimensional point cloud to generate an initial radiation field; wherein the initial radiation field includes a plurality of two-dimensional Gaussian points, each of which corresponds to a point cloud; Initialize the parameters of each two-dimensional Gaussian point, and adjust the parameters of each two-dimensional Gaussian point based on the differentiable rendering method to obtain the optimized radiation field.
3. The three-dimensional reconstruction method based on compressed signed distance field according to claim 1, characterized in that: The extracting parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding facet comprises: Extracting parameters of each two-dimensional Gaussian point from the radiation field to construct a corresponding initial facet; Each initial surface element is subjected to perspective correction and rendering processing to obtain the final surface element.
4. The three-dimensional reconstruction method based on compressed signed distance field according to claim 3, characterized in that: The perspective correction and rendering processing are performed on each initial face element to obtain the final face element, including: Perform perspective projection and correction processing on each initial face element to obtain a corrected face element; Adjusting the coordinates of each corrected face element based on the ray-ellipse intersection method to obtain an adjusted face element; Performing anti-aliasing processing on each adjusted face element to obtain an anti-aliased face element; Each anti-aliased surface element is rasterized to obtain the final surface element.
5. The three-dimensional reconstruction method based on compressed signed distance field according to claim 1, characterized in that: The Gaussian bounding box optimization algorithm and the deep warping back propagation algorithm based on dynamic programming perform performance optimization and depth adjustment on each face element to obtain an optimized face element, including: Based on the opacity of each face element, the radius of the Gaussian bounding box of each face element is calculated; Initialize the depth and weight of each face element and calculate the depth loss of each face element; According to each depth loss, the depth and weight of the corresponding surface element are optimized to obtain the corresponding optimized surface element.
6. The three-dimensional reconstruction method based on compressed signed distance field according to claim 5, characterized in that: The radius of the Gaussian bounding box is Among them, α is the opacity and r is the radius of the Gaussian bounding box.
7. The three-dimensional reconstruction method based on compressed signed distance field according to claim 1, characterized in that: The method of extracting three-dimensional mesh models from all optimized facets based on the compressed signed distance field, and constructing corresponding three-dimensional scenes according to all extracted three-dimensional mesh models, includes: According to the depth of the preset grid points in different optimized facets, the truncated signed distance of the grid points is determined to generate a signed distance field; Extracting isosurfaces from the signed distance field using a marching cubes algorithm to obtain a three-dimensional mesh model corresponding to each optimized face element; All three-dimensional grid models are merged into an overall three-dimensional grid model, and a corresponding three-dimensional scene is constructed according to the overall three-dimensional grid model.
8. The three-dimensional reconstruction method based on compressed signed distance field according to claim 7, characterized in that: The method of determining the truncated signed distance of the grid points according to the depths of the preset grid points in the different optimized facets and generating the signed distance field comprises: Normalize all preset grid points; For each grid point: Map the grid points to the world coordinate system through inverse transformation; Determine the cutoff signed distance of the grid point according to the depth of the grid point in the different optimized bins; The truncated signed distance of each grid point is summed up to obtain the signed distance field.
9. A three-dimensional reconstruction system based on compressed signed distance field, characterized in that: The system comprises: A data acquisition module is used to obtain images of different viewing angles of the scene to be reconstructed and the camera poses of the corresponding viewing angles to generate a three-dimensional point cloud; A radiation field construction module, used to perform radiation field modeling and differentiable rendering processing on the three-dimensional point cloud to obtain a radiation field; wherein the radiation field includes a plurality of two-dimensional Gaussian points, each of which corresponds to a point cloud; A facet modeling module, used to extract parameters of each two-dimensional Gaussian point from the radiation field and construct corresponding facets; The optimization module is used to optimize the performance and adjust the depth of each face element based on the Gaussian bounding box optimization algorithm and the deep warping back propagation algorithm of dynamic programming to obtain the optimized face element; The reconstruction module is used to extract the three-dimensional mesh models from all the optimized face elements based on the compressed signed distance field, and construct the corresponding three-dimensional scene according to all the extracted three-dimensional mesh models.
10. An electronic device, characterized in that: The electronic device comprises: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, enables the electronic device to implement the three-dimensional reconstruction method based on the compressed signed distance field as described in any one of claims 1 to 8.
11. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor of a computer, the computer is enabled to execute the three-dimensional reconstruction method based on a compressed signed distance field as claimed in any one of claims 1 to 8.
Citation Information
Cited By
Foreground target segmentation and reconstruction method and device, equipment and storage medium
CN119942544A
Foreground object segmentation and reconstruction method, device, equipment and storage medium
CN119942544B
Multi-model collaborative two-dimensional Gaussian splash three-dimensional reconstruction method
CN121392157A
