Whole body three-dimensional human body mesh generation method and device, storage medium and computer equipment

Abstract

The method, apparatus, storage medium, and computer equipment for generating full-body 3D human body meshes provided in this application acquire a full-body human image and crop it to obtain an upper-body human image; generate a full-body source mesh and a half-body reference mesh corresponding to the full-body human image, wherein the half-body reference mesh has a higher mesh precision than the full-body source mesh; determine the effective cropping direction based on the cropping box boundary of the upper-body human image, and crop the full-body source mesh and the half-body reference mesh respectively based on the effective cropping direction; adjust the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh, and then use the adjusted pose as the basis for the final image. Using a half-body reference mesh as the target, non-rigid meshing is performed on the seams of the cropped full-body source mesh to obtain an initial 3D human body mesh. The initial 3D human body mesh is then fused, and the fused initial 3D human body mesh is further densified to obtain an intermediate 3D human body mesh. The intermediate 3D human body mesh is then non-rigidly fused again, and the fused intermediate 3D human body mesh is subjected to surface reduction and texture baking to obtain the final full-body 3D human body mesh. This method avoids the convergence instability or local mesh folding caused by direct non-rigid registration, effectively restores the local high-frequency details lost during the fusion process, and compensates for the loss of accuracy in a single fusion.

Description

Technical Field

[0001] This application relates to the fields of three-dimensional computer vision and 3D digital human body generation technology, and in particular to a method, device, storage medium and computer equipment for generating a full-body three-dimensional human body mesh. Background Technology

[0002] Currently, large-scale 3D human body generation models can reconstruct a complete full-body 3D human body mesh from a single full-body photograph. However, due to the limited voxel space resolution, the generated full-body mesh suffers from blurry details and inaccurate curvature representation in the upper body region (head, neck, shoulders, chest, etc.). While 3D generation models specifically trained for the half-body can reconstruct high-precision upper-body meshes, they cannot directly generate full-body models.

[0003] Existing technologies attempt to combine high-precision half-body meshes with full-body meshes, but they suffer from the following problems: global rigid alignment methods can only eliminate overall translation and rotation deviations and cannot handle local non-rigid deformations at the seams; global non-rigid registration methods lack spatial regional weight control, which can easily lead to excessive stretching and deformation of normal regions far from the seams; simple splicing can cause geometric breaks and normal jumps at the seams; the smoothing effect of single-pass fusion methods can sacrifice high-frequency details at the seams and cannot restore the original accuracy of the half-body mesh; and the above schemes can easily lead to unstable convergence or local mesh folding when performing non-rigid registration directly under large pose deviations. Summary of the Invention

[0004] The purpose of this application is to at least solve one of the above-mentioned technical defects, especially the technical defects in the prior art where the fusion accuracy at the joint is low when combining high-precision half-body mesh with full-body mesh, and it is easy to cause unstable convergence or local mesh folding.

[0005] This application provides a method for generating a full-body 3D human mesh, the method comprising:

[0006] A full-body image is acquired, and the full-body image is cropped to obtain an upper-body image.

[0007] Generate a full-body source mesh corresponding to the full-body human image and a half-body reference mesh corresponding to the upper-body human image, wherein the half-body reference mesh has a higher mesh precision than the full-body source mesh;

[0008] The effective cropping direction is determined based on the cropping box boundary of the upper body human image, and the full-body source mesh and the half-body reference mesh are cropped based on the effective cropping direction;

[0009] Based on the cropped full-body source mesh and the cropped half-body reference mesh, the pose of the half-body reference mesh is adjusted, and the mesh at the seam of the cropped full-body source mesh is non-rigidly snapped with the pose-adjusted half-body reference mesh as the target to obtain the initial three-dimensional human body mesh.

[0010] The initial three-dimensional human body mesh is fused, and the fused initial three-dimensional human body mesh is densified to obtain an intermediate three-dimensional human body mesh.

[0011] The intermediate three-dimensional human body mesh is then subjected to non-rigid adsorption again, and the adsorbed intermediate three-dimensional human body mesh is subjected to surface reduction and texture baking to obtain the final full-body three-dimensional human body mesh.

[0012] Optionally, determining the effective cropping direction based on the cropping frame boundary of the upper body image includes:

[0013] Determine whether there are valid pixel values ​​at the top, bottom, left, and right boundaries of the cropping frame of the upper body image;

[0014] The boundary with valid pixel values ​​is determined as the valid boundary, and the valid boundary is mapped to the effective cropping direction in three-dimensional space according to the imaging viewpoint parameters of the whole body image.

[0015] Optionally, the effective clipping direction is configured using a clipping surface data structure, with each effective clipping direction corresponding to a clipping surface;

[0016] The cutting surface includes four parameters: arrangement axis information, scanning axis, scanning direction, and band length ratio.

[0017] The arrangement axis information is composed of the arrangement order of the other two coordinate axes besides the scanning axis among the three coordinate axes x, y, and z. The scanning axis is the x-axis, y-axis, or z-axis. The scanning direction is the positive or negative direction along the scanning axis. The band length ratio is the proportion of the clipping surface along the scanning axis direction to the span of the target mesh bounding box, used to determine the end position of the clipping surface along the scanning axis direction.

[0018] Optionally, the step of cropping the full-body source mesh and the half-body reference mesh based on the effective cropping direction includes:

[0019] For the clipping surface corresponding to the effective clipping direction, the set of vertices that extend beyond the endpoint of the clipping surface along the corresponding scan direction coordinates is taken as the out-of-band far-end region of the clipping surface; wherein, when there are multiple effective clipping directions, the out-of-band far-end region is the intersection of the out-of-band far-end regions of the clipping surfaces corresponding to each effective clipping direction.

[0020] When trimming the full-body source mesh, delete the vertices and associated triangles belonging to the out-of-band far-end region in the full-body source mesh, and retain the area near the seam and the main body;

[0021] When clipping the half-body reference mesh, delete the vertices in the half-body reference mesh that do not belong to the out-of-band far-end region, and only retain the mesh portion in the half-body reference mesh that belongs to the out-of-band far-end region.

[0022] Optionally, adjusting the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh includes:

[0023] Based on the axis-aligned bounding box of the cropped half-body reference mesh, extract the vertices that fall within the outward expansion range of the axis-aligned bounding box from the cropped full-body source mesh as the first point set, and extract the vertices that fall outside the inward contraction range of the axis-aligned bounding box from the cropped half-body reference mesh as the second point set.

[0024] Determine whether the number of vertices in the second point set has reached a preset threshold.

[0025] If so, the rigid transformation matrix between the first point set and the second point set is estimated by the rigid registration algorithm, and the rigid transformation matrix is ​​applied to the half-body reference mesh to adjust the pose of the half-body reference mesh;

[0026] Otherwise, the half-body reference mesh is kept in its original pose.

[0027] Optionally, the step of non-rigidly snapping the mesh at the seams of the cropped full-body source mesh, using the pose-adjusted half-body reference mesh as the target, includes:

[0028] Within the seam zone of the cropped full-body source mesh, the gradient weight of each vertex is calculated, and the gradient weight decreases monotonically from the seam towards the inside of the seam zone.

[0029] Using the pose-adjusted half-body reference mesh as the target, the vertices in the seam area are subjected to non-rigid deformation according to the gradient weight, so that the adsorption strength of the vertices near the seam is greater than that of the vertices far from the seam.

[0030] Optionally, calculating the gradient weights of each vertex within the seam region of the cropped full-body source mesh includes:

[0031] Determine the cutting surface corresponding to the effective cutting direction;

[0032] Within the seam zone of the trimmed full-body source mesh, vertices that meet the eligibility criteria are determined based on the trimming surfaces. The eligibility criteria include: the vertex is not excluded from the far-end region outside the strip by any of the trimming surfaces, and falls at least within the strip zone of any of the trimming surfaces.

[0033] For each vertex, when the vertex falls within the in-band interval of multiple clipping planes, the weight value of the vertex in each clipping plane is calculated based on the position of the vertex in the in-band interval of each clipping plane, and the minimum value among the weight values ​​is taken as the gradient weight of the vertex.

[0034] When a vertex falls within the band of only one clipping plane, calculate the weight value of the vertex in that clipping plane and use the calculated weight value as the gradient weight of the vertex.

[0035] Optionally, calculating the weight value of the vertex in the clipping plane includes:

[0036] The intra-band interval of the clipping surface is divided into several sub-bands along the scanning direction, and the sub-band index to which the vertex belongs in the intra-band interval is determined;

[0037] The weight value of the vertex in the cutting plane is obtained by normalizing the relative position of the sub-band index in the interval within the band, and the weight value decreases from the outside closer to the seam surface to the inside farther away from the seam surface.

[0038] Optionally, the step of using the pose-adjusted half-body reference mesh as the target and applying non-rigid deformation to the vertices within the seam area according to the gradient weights includes:

[0039] A local affine deformation model is constructed, and local affine transformation parameters are defined on the mesh edges of the seam zone. Deformation is transferred through graph propagation.

[0040] With the goal of minimizing the energy function, the local affine transformation parameters of each mesh edge in the joint zone are solved by optimizing the energy function, and the vertices in the joint zone are subjected to non-rigid deformation based on the solved local affine transformation parameters.

[0041] The energy function includes a data term, a preservation term, a stiffness term, and a Laplacian smoothing term. The data term is used to pull the vertex to the corresponding target position on the pose-adjusted half-body reference mesh, and the weight of the data term is positively correlated with the gradient weight. The preservation term is used to constrain the deformed vertex to its original position, and the weight of the preservation term is negatively correlated with the gradient weight. The stiffness term is used to constrain the local affine transformation parameters. The Laplacian smoothing term is used to preserve the local smoothness of the deformed mesh.

[0042] Optionally, the process of determining the target location includes:

[0043] For each vertex in the seam area, take K nearest neighbor vertices on the pose-adjusted half-body reference mesh, and obtain the K nearest neighbor soft anchor point positions by distance-weighted average, and use the K nearest neighbor soft anchor point positions as the target positions;

[0044] Where K is a positive integer greater than or equal to 2.

[0045] Optionally, the mesh fusion of the initial three-dimensional human body mesh includes:

[0046] The upper body part after non-rigid adsorption is merged with the lower body part of the whole body source mesh into a single mesh, and the single mesh is reconstructed at the seam to make the vertices and triangular faces on both sides of the seam topologically continuous.

[0047] The reconstruction process can be implemented in any one of the following ways: Poisson fusion, edge-stitched mesh fusion, or implicit surface-based mesh fusion.

[0048] Optionally, the non-rigid adsorption of the intermediate three-dimensional human body mesh again includes:

[0049] Within the seam zone of the intermediate three-dimensional human body mesh, a constant adsorption weight is assigned to each vertex.

[0050] Using the half-body reference mesh as a high-precision geometric target, non-rigid deformation is applied to the vertices within the seam area according to the adsorption weight;

[0051] The stiffness constraint strength during the second non-rigid adsorption is less than that during the first non-rigid adsorption.

[0052] Optionally, the step of performing surface reduction and texture baking on the adsorbed intermediate 3D human body mesh to obtain the final full-body 3D human body mesh includes:

[0053] The quadratic error metric method is used to reduce the number of triangular faces in the intermediate three-dimensional human body mesh after adsorption, thereby reducing the number of triangular faces to the target size.

[0054] The intermediate 3D human body mesh after surface reduction is automatically unwrapped and texture maps are baked to generate a renderable 3D asset format, resulting in the final full-body 3D human body mesh.

[0055] This application also provides a whole-body three-dimensional human mesh generation device, including:

[0056] The image cropping module is used to acquire a full-body human image and crop the full-body human image to obtain an upper-body human image.

[0057] The mesh generation module is used to generate a full-body source mesh corresponding to the full-body human image and a half-body reference mesh corresponding to the upper-body human image, wherein the mesh accuracy of the half-body reference mesh is higher than that of the full-body source mesh.

[0058] The mesh cropping module is used to determine the effective cropping direction based on the cropping box boundary of the upper body human image, and to crop the full-body source mesh and the half-body reference mesh based on the effective cropping direction.

[0059] The pose adjustment and adsorption module is used to adjust the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh, and to non-rigidly adsorb the mesh at the seam of the cropped full-body source mesh with the pose-adjusted half-body reference mesh as the target to obtain the initial three-dimensional human body mesh.

[0060] The mesh fusion and densification module is used to perform mesh fusion on the initial three-dimensional human body mesh and to densify the fused initial three-dimensional human body mesh to obtain an intermediate three-dimensional human body mesh.

[0061] The mesh reduction and baking module is used to perform non-rigid adsorption on the intermediate three-dimensional human body mesh again, and to perform surface reduction and texture baking on the adsorbed intermediate three-dimensional human body mesh to obtain the final full-body three-dimensional human body mesh.

[0062] This application also provides a computer-readable storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the full-body three-dimensional human mesh generation method as described in any of the above embodiments.

[0063] This application also provides a computer device, including: one or more processors, and memory;

[0064] The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the full-body three-dimensional human mesh generation method as described in any of the above embodiments.

[0065] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0066] The method, apparatus, storage medium, and computer equipment for generating full-body 3D human meshes provided in this application employ a technical route of "trimming and positioning → pose adjustment → non-rigid adsorption → fusion and densification → secondary adsorption → post-processing." Using a high-precision half-body reference mesh as the target, non-rigid adsorption is applied to the seams of the full-body source mesh, ensuring that the accuracy of the upper half-body region of the full-body mesh is comparable to that of the half-body reference mesh. In the pose adjustment step, rigid registration is performed by extracting shell point clouds from the trimmed half-body reference mesh and the full-body source mesh. Only the half-body reference mesh is subjected to rigid transformation, while the full-body source mesh remains stationary, even if there is a significant difference in position between the half-body reference mesh and the full-body source mesh. Even with pose deviations, the half-body reference mesh can be stably aligned to the full-body coordinate system, providing good initial conditions for subsequent non-rigid registration and avoiding convergence instability or local mesh folding caused by direct non-rigid registration. Next, this application ensures the overall topological continuity of the full-body mesh through mesh fusion, eliminating seam traces. Then, a second non-rigid registration is performed using the half-body reference mesh as a high-precision geometric target, effectively restoring the local high-frequency details lost during the fusion process and compensating for the accuracy loss of a single fusion. Finally, after face reduction and texture baking, a renderable 3D asset with a moderate number of triangles and UV texture mapping is output, meeting the direct use requirements of downstream applications. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 A flowchart illustrating a method for generating a full-body three-dimensional human mesh, as provided in an embodiment of this application;

[0069] Figure 2 This is a structural illustration of the human body image and various cutting surfaces provided in the embodiments of this application;

[0070] Figure 3 This is a schematic diagram illustrating the process of determining whether to adjust the pose of the half-body reference mesh in an embodiment of this application;

[0071] Figure 4 This is a schematic diagram of the structure of a three-dimensional human body mesh generation device provided in an embodiment of this application;

[0072] Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0073] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0074] Currently, existing 3D human body mesh generation technologies, when dealing with full-body reconstruction scenarios, are often limited by image resolution, video memory capacity, and algorithm complexity, making it difficult to directly generate full-body 3D meshes that combine global integrity and local high precision. If the entire full-body image is used as input to generate the mesh, the accuracy of areas that need to retain details, such as the upper body, is often insufficient due to limitations in model computation, failing to meet the requirements for restoring human posture and clothing details. If the full-body image is directly cropped into the upper body region for separate reconstruction, only a half-body mesh can be obtained, and a complete full-body mesh asset cannot be directly obtained. Furthermore, there is a pose deviation between the separately reconstructed half-body mesh and the directly generated full-body mesh, and direct fusion is prone to problems such as topological discontinuity, loss of details, and deformation misalignment.

[0075] This application addresses these pain points in existing technologies through the following technical solutions, achieving both the integrity of the full-body mesh and high-precision detail restoration of the upper body region while controlling the overall computational load, as detailed below:

[0076] In one embodiment, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for generating a full-body 3D human body mesh, as provided in an embodiment of this application. This application provides a method for generating a full-body 3D human body mesh, which may include:

[0077] S110: Obtain a full-body image and crop it to obtain an upper-body image.

[0078] In this step, when generating a full-body 3D human body mesh, this application can first detect the human body region in the input full-body human body image using a preset detection model, estimate the upper body range based on the human body posture, and automatically generate the corresponding upper body region cropping box. Alternatively, it can support manual annotation to determine the cropping box position, and directly complete the cropping based on the cropping box boundary to obtain the upper body human body image.

[0079] In this application, the full-body human image refers to a human image that includes the complete human torso and limbs, while the upper-body human image is a partial human image cropped from the full-body human image, containing only the upper half of the human body, such as the head, neck, shoulders, and chest. The specific human body area included can be determined according to the actual situation and is not limited here. Compared to the full-body human image, the cropped upper-body human image has a higher resolution and can provide sufficient detail information for the subsequent generation of a high-precision half-body reference mesh.

[0080] S120: Generate a full-body source mesh corresponding to the full-body human image, and a half-body reference mesh corresponding to the upper-body human image. The half-body reference mesh has a higher mesh accuracy than the full-body source mesh.

[0081] In this step, a full-body image is acquired via S110, and after cropping the full-body image to obtain an upper-body image, the corresponding initial 3D human body meshes can be output by calling a preset 3D human body reconstruction model, using the full-body image and the upper-body image as input respectively. The result obtained by using the full-body image as input is the full-body source mesh, which contains the complete structure of the human body. Due to the computational load, the overall mesh accuracy is relatively low. The result obtained by using the upper-body image as input is the half-body reference mesh, which only contains the structure of the upper body. However, because the local resolution of the input image is higher and the model can focus on calculating the upper body region, the mesh accuracy is significantly higher than that of the full-body source mesh, and it can restore richer details of clothing and posture.

[0082] S130: Determine the effective cropping direction based on the cropping box boundary of the upper body human image, and crop the full-body source mesh and the half-body reference mesh based on the effective cropping direction.

[0083] In this step, after obtaining two initial meshes with different precisions through S120, the half-body reference mesh is first projected onto the unified three-dimensional coordinate system where the full-body source mesh is located through the correspondence in image space. Then, the corresponding effective clipping direction is determined in three-dimensional space according to the original clipping box boundary. The full-body source mesh and the half-body reference mesh are clipped along the effective clipping direction: the original low-precision part of the upper body region corresponding to the full-body source mesh is clipped, and only the complete lower body and the edge seam area of ​​the upper body are retained; the excess part of the half-body reference mesh that exceeds the clipping range is clipped, and only the high-precision upper body region is retained, thus completing the preprocessing of the two meshes.

[0084] In this application, the effective cropping direction refers to the scanning direction parallel to any of the top, bottom, left, and right boundaries of the upper body image cropping frame. Scanning all vertices along this direction allows determination of whether each vertex needs to be retained and whether it is located within the seam area required for subsequent deformation, based on its coordinates. During the scanning process, this application can reserve a pre-defined width interval based on the retention boundary as the seam area within the band, used for subsequent non-rigid deformation fusion, thus avoiding obvious splicing marks at the seams.

[0085] S140: Based on the cropped full-body source mesh and the cropped half-body reference mesh, adjust the pose of the half-body reference mesh, and use the pose-adjusted half-body reference mesh as the target to perform non-rigid adsorption on the mesh at the seam of the cropped full-body source mesh to obtain the initial three-dimensional human body mesh.

[0086] In this step, after the clipping operation is completed through S130, since there is an initial pose deviation between the half-body reference mesh obtained by separate reconstruction and the whole-body source mesh, it is necessary to first adjust the pose of the half-body reference mesh through rigid registration and align it to the coordinate system of the whole-body source mesh, and then carry out the subsequent non-rigid adsorption operation based on the alignment result.

[0087] Specifically, this step first extracts shell point clouds from the overlapping regions of the cropped half-body reference mesh and the full-body source mesh. Rigid registration is then performed on the two point clouds to obtain rigid transformation parameters. These transformation parameters are applied to the entire half-body reference mesh to complete pose adjustment, aligning the half-body reference mesh to the corresponding region of the full-body source mesh, while the full-body source mesh remains stationary throughout this process. After pose adjustment, within the seam region retained in the cropped full-body source mesh, weights that gradually change with relative position are assigned to each vertex. Then, using the pose-adjusted half-body reference mesh as the target, non-rigid deformation is performed on the vertices within the seam region based on the gradual weights. The seam vertices are gradually snapped to their corresponding high-precision geometric positions within the pose-adjusted half-body reference mesh, resulting in the initial 3D human body mesh before fusion.

[0088] S150: Perform mesh fusion on the initial 3D human body mesh, and then perform mesh densification on the fused initial 3D human body mesh to obtain an intermediate 3D human body mesh.

[0089] In this step, after obtaining the initial three-dimensional human body mesh through non-rigid adsorption via S140, it is necessary to integrate the upper body part after non-rigid adsorption with the lower body part retained by the whole body source mesh into a continuous and complete overall mesh.

[0090] Specifically, this application first merges two meshes into a single mesh, then performs topological reconstruction on the seam areas of the single mesh to eliminate topological breaks on both sides of the seam, ensuring continuous connection between the vertices and triangular faces of the entire mesh and avoiding gaps or overlapping faces. After mesh fusion, in order to match the high-precision details of the half-body reference mesh, this application can perform overall mesh densification on the currently fused complete mesh, refining the triangular faces of the low-precision mesh, improving the overall resolution of the entire mesh, and obtaining an intermediate 3D human body mesh that meets the detail-carrying capacity.

[0091] In one specific implementation, when densifying the overall mesh, this application can adopt a Loop-based subdivision densification method to sequentially subdivide all triangular faces, thereby increasing the number of mesh vertices and face density without changing the overall shape, providing a sufficient topological foundation for preserving high-precision geometric details in the future.

[0092] S160: The intermediate 3D human body mesh is non-rigidly adsorbed again, and the adsorbed intermediate 3D human body mesh is reduced in surface area and texture baked to obtain the final full-body 3D human body mesh.

[0093] In this step, after obtaining the overall densified intermediate 3D human body mesh through S150, a second non-rigid snapping is performed on the upper half of the intermediate 3D human body mesh, using a high-precision half-body reference mesh as the geometric target. This snaps the vertices of the entire upper half of the body directly to the corresponding geometric positions on the half-body reference mesh, restoring the local high-frequency details lost during mesh fusion and overall densification, while ensuring the mesh as a whole maintains a continuous and complete topological structure, avoiding the loss of local details. During the snapping process, the smoothness of the snapping deformation can be adjusted through preset curvature constraints to avoid excessive offset of local vertices, resulting in mesh folding or unevenness.

[0094] Understandably, the overall accuracy of the full-body 3D mesh obtained after two non-rigid adsorption processes already meets the requirements for detail reproduction. However, the overall triangle count of the mesh after densification is too high, making it unsuitable for direct rendering in downstream applications. Therefore, this application requires facet reduction processing of the entire mesh to control the overall triangle count while preserving geometric details as much as possible, resulting in a final mesh with a moderate facet count and uniform topology. After facet reduction, the color information of the original input image is used to generate corresponding UV texture maps through texture baking, ultimately outputting a full-body 3D human body mesh asset with complete texture.

[0095] In the above embodiments, the technical route of "cropping and positioning → pose adjustment → non-rigid adsorption → fusion and densification → secondary adsorption → post-processing" is adopted. A high-precision half-body reference mesh is used as the target to perform non-rigid adsorption at the seams of the full-body source mesh, making the accuracy of the upper half of the full-body mesh comparable to that of the half-body reference mesh. In the pose adjustment step, the shell point cloud is extracted from the cropped half-body reference mesh and the full-body source mesh for rigid registration. Only the half-body reference mesh is subjected to rigid transformation, while the full-body source mesh remains stationary. Even if there is a large pose deviation between the half-body reference mesh and the full-body source mesh, the half-body reference mesh can be stably registered. Aligning the reference mesh to the whole-body coordinate system provides a good initial condition for subsequent non-rigid registration, avoiding convergence instability or local mesh folding caused by direct non-rigid registration. Next, this application ensures the overall topological continuity of the whole-body mesh through mesh fusion, eliminating seam traces. Then, a second non-rigid registration is performed using the half-body reference mesh as a high-precision geometric target, effectively restoring the local high-frequency details lost during the fusion process and compensating for the accuracy loss of a single fusion. Finally, after face reduction and texture baking, a renderable 3D asset with a moderate number of triangles and UV texture mapping is output, meeting the direct use requirements of downstream applications.

[0096] In one embodiment, determining the effective cropping direction based on the cropping frame boundary of the upper body image in step S130 may include:

[0097] S131: Determine whether there are valid pixel values ​​at the top, bottom, left, and right boundaries of the cropping frame of the upper body image.

[0098] S132: Determine the boundary with valid pixel values ​​as the valid boundary, and map the valid boundary to the effective cropping direction in three-dimensional space according to the imaging perspective parameters of the whole body image.

[0099] In this embodiment, when determining the effective cropping direction, this application can first judge each boundary of the cropping box one by one. If there are valid pixels belonging to the human body area on a certain boundary, it means that the boundary is a valid boundary that needs to retain the human body content and needs to complete the vertex scanning judgment along the direction. Otherwise, it means that the boundary coincides with the image edge and there is no need to set the cropping direction separately.

[0100] Subsequently, this application can transform all effective boundaries from image space to three-dimensional space according to the imaging projection relationship to obtain the effective clipping direction in the corresponding direction. Then, vertex scanning can be completed sequentially along each effective clipping direction to distinguish the mesh area to be retained from the mesh area to be clipped, and a seam strip area of ​​preset width is reserved at the edge of the retained area to ensure sufficient adjustment space for subsequent deformation fusion.

[0101] In one embodiment, the effective clipping direction is configured using a clipping surface data structure, with each effective clipping direction corresponding to a clipping surface.

[0102] The cutting surface includes four parameters: arrangement axis information, scanning axis, scanning direction, and band length ratio.

[0103] The arrangement axis information is composed of the arrangement order of the other two coordinate axes besides the scanning axis among the three coordinate axes x, y, and z. The scanning axis is the x-axis, y-axis, or z-axis. The scanning direction is the positive or negative direction along the scanning axis. The band length ratio is the proportion of the clipping surface along the scanning axis direction to the span of the target mesh bounding box, used to determine the end position of the clipping surface along the scanning axis direction.

[0104] In this embodiment, each effective clipping direction corresponds to a clipping plane, and each clipping plane corresponds to an effective boundary in the image space. Four parameters precisely describe the clipping range and seam allowance rules in that direction: the scanning axis corresponds to the direction perpendicular to the effective boundary in 3D space; the scanning direction determines which direction to start scanning the mesh vertices; the arrangement axis information is used to organize the scanning order of all vertices; and the seam allowance ratio is automatically calculated based on the overall size of the target mesh to determine the required seam allowance width, eliminating the need for manual input of absolute dimensions and adapting to human body meshes of different sizes. When scanning mesh vertices, simply scan sequentially according to the clipping plane parameters to quickly distinguish the target area to be retained, the excess area to be clipped, and the seam allowance area for subsequent deformation fusion. The calculation logic is clear and stable, adapting to different clipping positions and different human body postures.

[0105] Indicatively, such as Figure 2 As shown, Figure 2 This is a structural illustration of the human body image and various clipping surfaces provided in the embodiments of this application. The clipping surfaces in this application are dictionary structures, with keys being 4-character strings (such as "yzx+", "xyz-", etc.). The rules are as follows: the first three characters are an arrangement of x, y, and z; the third character is the scanning axis; the fourth character is "+" or "-", indicating the scanning direction along the scanning axis; and the value is a floating-point number in (0, 1), representing the proportion of the clipping surface along the scanning axis to the span of the target mesh bounding box.

[0106] Furthermore, for clipping planes marked with a "+", the starting point is the minimum coordinate of the target mesh on that axis, and the ending point is the minimum coordinate plus the scale multiplied by the span; for clipping planes marked with a "-", the starting point is the maximum coordinate, and the ending point is the maximum coordinate minus the scale multiplied by the span. The span is the difference between the maximum and minimum coordinates.

[0107] The clipping face switch is a 0 / 1 list corresponding to the clipping face key sequence. For example, [1, 1, 1, 0] indicates that the first three clipping faces are enabled and the fourth is disabled. Only items with a switch of 1 participate in subsequent weight calculations and mesh clipping. This mechanism is used to manually specify which axes have real seams in the actual human body, avoiding misalignment of natural boundaries without cuts.

[0108] In the above embodiments, by configuring a cutting surface data structure containing four parameters—arrangement axis information, scanning axis, scanning direction, and strip length ratio—for each effective cutting direction, precise and automatic adaptation between the cutting range and the reserved seam tape is achieved. There is no need to manually input the absolute size of the seam tape. The reserved width can be automatically calculated based on human body meshes of different sizes. At the same time, the effective cutting direction is controlled by the cutting surface switch, avoiding misoperation on natural boundaries without cuts. It adapts to application scenarios with different cutting positions and different human body postures, and the cutting calculation logic is clear and stable.

[0109] In one embodiment, S130, which involves clipping the full-body source mesh and the half-body reference mesh based on the effective clipping direction, may include:

[0110] S133: For the clipping surface corresponding to the effective clipping direction, the set of vertices that extend beyond the endpoint of the clipping surface along the corresponding scanning direction coordinates is taken as the out-of-band far-end region of the clipping surface; wherein, when there are multiple effective clipping directions, the out-of-band far-end region is the intersection of the out-of-band far-end regions of the clipping surfaces corresponding to each effective clipping direction.

[0111] S134: When trimming the whole-body source mesh, delete the vertices and associated triangular faces of the out-of-band far-end region in the whole-body source mesh, and retain the area near the seam and the main body.

[0112] S135: When clipping the half-body reference mesh, delete the vertices in the half-body reference mesh that do not belong to the out-of-band far-end region, and only retain the mesh portion in the half-body reference mesh that belongs to the out-of-band far-end region.

[0113] In this embodiment, when trimming the full-body source mesh and the half-body reference mesh respectively, the regions to be deleted for the two meshes are exactly opposite according to the definition of the out-of-band far-end region: For the full-body source mesh, the out-of-band far-end region is the part of the upper body that is outside the trimming range and originally had lower precision and needs to be replaced. Therefore, after deleting this region, only the lower body and the seam area at the edge of the upper body can be retained; for the half-body reference mesh, only the high-precision target upper body region needs to be retained, and the lower body part is exactly outside the out-of-band far-end region. Therefore, after deleting the region that does not belong to the out-of-band far-end, only the high-precision mesh part of the upper body that meets the requirements can be retained. This opposite deletion rule naturally adapts to the trimming needs of the two meshes, without the need to set additional differential judgment logic. The corresponding trimming of the two meshes can be completed simultaneously with only one set of trimming surface parameters, simplifying the calculation process and ensuring the matching of the trimming regions.

[0114] When there is only one effective clipping direction, the out-of-band far-end region of this application is the out-of-band far-end region of the clipping surface corresponding to the effective clipping direction itself; when there are multiple effective clipping directions, the out-of-band far-end region is the intersection of the out-of-band far-end regions obtained from each clipping surface, and only the vertices that simultaneously meet the requirements of all clipping surface regions are retained, accurately matching the requirement of clipping in multiple directions at the same time, and obtaining the clipping regions corresponding to the two meshes, providing an accurate regional basis for subsequent pose adjustment and non-rigid adsorption.

[0115] This opposite region deletion rule ensures accurate matching of the two mesh clipping regions while reusing the same set of clipping surface parameters. It eliminates the need to design separate clipping logic for the full-body source mesh and the half-body reference mesh, greatly simplifying the calculation process. It also avoids the problem of misalignment between the two mesh clipping regions from the rule level, providing a reliable regional basis for subsequent pose registration and deformation fusion.

[0116] In one embodiment, such as Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the process of determining whether to adjust the pose of the half-body reference mesh in an embodiment of this application; S140, adjusting the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh, may include:

[0117] S141: Using the axis-aligned bounding box of the cropped half-body reference mesh as a reference, extract vertices from the cropped full-body source mesh that fall within the outward expansion range of the axis-aligned bounding box as a first point set, and extract vertices from the cropped half-body reference mesh that fall outside the inward contraction range of the axis-aligned bounding box as a second point set.

[0118] S142: Determine whether the number of vertices in the second point set has reached a preset threshold.

[0119] S143: If so, the rigid transformation matrix between the first point set and the second point set is estimated by the rigid registration algorithm, and the rigid transformation matrix is ​​applied to the half-body reference mesh to adjust the pose of the half-body reference mesh.

[0120] S144: Otherwise, keep the half-body reference mesh in its original pose.

[0121] In this embodiment, when adjusting the pose of the half-body reference mesh, corresponding point sets are first extracted from two meshes through bounding box expansion and contraction: the first point set is taken from the vertices of the full-body source mesh within the bounding box expansion range, corresponding to the area that the half-body reference mesh should originally cover; the second point set is taken from the vertices of the half-body reference mesh outside the bounding box contraction range, corresponding to the effective main body area of ​​the half-body reference mesh. If the number of vertices in the second point set reaches a preset threshold, it means that the current half-body reference mesh has sufficient overlap and matching basis with the full-body source mesh area, and the rigid transformation matrix can be directly estimated to complete the pose alignment; otherwise, it means that the overlap between the half-body reference mesh and the full-body source mesh is too low, and the extracted effective point set is insufficient to support stable rigid transformation estimation. In this case, the original pose of the half-body reference mesh is retained to avoid erroneous transformations that could lead to complete failure of subsequent registration.

[0122] The axis-aligned bounding box of the cropped half-body reference mesh refers to a bounding box parallel to the coordinate axes, formed by taking the maximum and minimum values ​​of the x, y, and z coordinates of all vertices of the cropped half-body reference mesh. This bounding box is used to quickly define the spatial area occupied by the half-body reference mesh, providing a unified range benchmark for the extraction of point sets for the subsequent two meshes. When extracting the first point set, this bounding box is expanded outward by a preset pixel distance along each of the three coordinate axes. All vertices of the full-body source mesh whose vertex coordinates fall within the expanded range are included in the first point set, ensuring coverage of the full-body area corresponding to the half-body reference mesh. When extracting the second point set, the original bounding box is shrunk inward by a preset pixel distance along each of the three coordinate axes. All vertices of the half-body reference mesh whose vertex coordinates are outside the shrunk range are included in the second point set, filtering out invalid matches where only a few vertices exist in the seam area, ensuring the reliability of the point set matching.

[0123] Furthermore, the aforementioned outward expansion and inward contraction preset pixel distances are configurable preset parameters that can be adaptively adjusted according to the overall scale of the human body mesh, ensuring that reliable corresponding point sets can be extracted from meshes of different scales. The aforementioned preset threshold can be preset according to the accuracy requirements of the actual application scenario, for example, set to XX% of the total number of vertices. Rigid transformation is only performed when the threshold condition is met; otherwise, pose adjustment is skipped. This adaptive judgment mechanism effectively avoids the problem of rigid registration errors leading to subsequent failures in the entire process due to insufficient overlap, improving the robustness of the entire algorithm to different input deviations.

[0124] The above embodiments automatically determine whether pose adjustment is needed by using a preset threshold number of values. This adapts to initial input scenarios with different degrees of overlap, avoids additional pose errors introduced by invalid transformations, and improves the robustness of the entire pose adjustment process. Even if the initial pose deviation is large, as long as there are enough effective overlapping vertices, rigid registration and alignment can be automatically triggered through this mechanism to eliminate the pose deviation. If the overlap is indeed insufficient and a reliable transformation cannot be estimated, the adjustment can be actively stopped to avoid introducing erroneous transformations, significantly improving the robustness of the pose adjustment steps and adapting to alignment requirements under different initial input conditions.

[0125] In one embodiment, S140, targeting the pose-adjusted half-body reference mesh, and performing non-rigid snapping on the seam mesh of the trimmed full-body source mesh, may include:

[0126] S145: Within the seam zone of the trimmed full-body source mesh, calculate the gradient weight of each vertex, wherein the gradient weight monotonically decreases from the seam towards the interior of the seam zone.

[0127] S146: Using the pose-adjusted half-body reference mesh as the target, perform non-rigid deformation on the vertices in the seam area according to the gradient weight, so that the adsorption strength of the vertices near the seam is greater than that of the vertices far from the seam.

[0128] In this embodiment, when performing non-rigid adsorption on the seam mesh of the cut full-body source mesh, this application calculates gradient weights within the seam area, allowing vertices closer to the seam to receive greater adsorption weights, thus more strongly conforming to the corresponding positions of the half-body reference mesh, while vertices farther from the seam have smaller weights and smaller deformation amplitudes, preserving the original shape of the full-body source mesh. This achieves a natural and smooth deformation transition at the seam between the two meshes, avoiding the abrupt misalignment caused by direct splicing.

[0129] Furthermore, the monotonically decreasing gradient weights along the seam in this application conform to the natural logic of deformation fusion. This ensures that the vertex positions of the two meshes at the seam are fully aligned, while avoiding unnecessary deformation effects on the main body of the source mesh far from the seam, thus preserving the effective geometric information of each mesh to the greatest extent. Simultaneously, controlling the adsorption intensity of non-rigid deformation with gradient weights eliminates the need for complex transition smoothing algorithms; a smooth transition can be achieved through the natural change in weights. This simplifies the deformation fusion calculation process and ensures a natural and continuous vertex distribution in the fused mesh, without obvious splicing marks.

[0130] In one specific implementation, the gradient weights in this application can be linear gradient weights corresponding to a strip weight curve, or other smoothly decreasing nonlinear weight curves, such as square root curves, cubic curves, etc., which can be selected and configured according to the actual fusion effect. Linear gradient weights have simple calculation logic and higher computational efficiency, while nonlinear gradient weights can achieve a smoother transition effect and adapt to scenarios with different precision requirements. By controlling the adsorption intensity through gradient weights, controllable deformation is generated only in the seam area, which not only completes the vertex alignment at the seam of the two meshes, but also preserves the original high-precision geometric information of the two meshes to the greatest extent, without generating additional deformation distortion, and finally obtains a natural and smooth full-body 3D human body fusion mesh.

[0131] In one embodiment, calculating the gradient weights of each vertex within the seam zone region of the trimmed full-body source mesh in step S145 may include:

[0132] S1451: Determine the cutting surface corresponding to the effective cutting direction.

[0133] S1452: Within the seam zone of the trimmed full-body source mesh, based on the trimming face, determine vertices that meet the qualification criteria, the qualification criteria including: the vertex is not excluded from the far-end region outside the band by any of the trimming faces, and falls at least within the band zone of any of the trimming faces.

[0134] S1453: For each vertex, when the vertex falls within the in-band interval of multiple clipping planes, calculate the weight value of the vertex in each clipping plane based on the vertex's position in the in-band interval of each clipping plane, and take the minimum value among the weight values ​​as the gradient weight of the vertex.

[0135] S1454: When the vertex falls within the band of only one clipping plane, calculate the weight value of the vertex in that clipping plane and use the calculated weight value as the gradient weight of the vertex.

[0136] In this embodiment, when calculating the gradient weights of vertices in the seam zone region of the whole-body source mesh, this application can first perform qualification screening on vertices based on each effective clipping face, retaining only valid vertices belonging to the seam zone region for weight calculation; for vertices that fall within multiple clipping face regions simultaneously, this application can take the minimum value among the multiple clipping face weight calculation results as the final gradient weight, ensuring that the adsorption strength of the vertex always meets the transition requirements of multiple clipping directions, avoiding unnatural transitions caused by excessive weight in a single direction. This rule is suitable for both single-direction clipping and multi-direction simultaneous clipping scenarios, with unified calculation logic, no need to set branch rules for different clipping numbers, simplifying the calculation process, and ensuring the rationality of the weight of each vertex in multi-direction clipping scenarios, providing an accurate weight basis for subsequent smooth deformation fusion.

[0137] For example, when a vertex lies within the band of both the x-axis and y-axis clipping planes, and its weight is calculated to be 0.8 in the x-axis direction and 0.5 in the y-axis direction, then 0.5 is taken as the final gradient weight for that vertex. This ensures that the vertex maintains the required deformation amplitude in both directions, preventing abrupt deformation transitions due to an excessively high weight in one direction. This rule of minimizing multi-directional weights naturally adapts to scenarios with multiple clipping directions, eliminating the need for complex multi-directional weight fusion calculations to guarantee smooth transitions in all clipping directions. The logic is simple and the effect is stable.

[0138] In one embodiment, calculating the weight value of the vertex in the clipping plane in step S1454 may include:

[0139] S14541: Divide the intra-band interval of the clipping surface into several sub-bands along the scanning direction, and determine the sub-band index to which the vertex belongs in the intra-band interval.

[0140] S14542: Normalize the relative position of the sub-band index in the band interval to obtain the weight value of the vertex in the cutting surface, wherein the weight value decreases from the outside near the seam surface to the inside away from the seam surface.

[0141] In this embodiment, since the in-band region itself is a strip-shaped area extending from the clipping surface towards the main body of the whole-body source mesh, the closer to the clipping surface (seam surface), the stronger the snapping alignment is required, while the further inward into the main body of the whole-body source mesh, the more the original shape needs to be preserved. Therefore, after normalizing the weights based on the relative positions of the vertex sub-band indices, the requirement of decreasing weights from the seam surface inward is naturally satisfied. The calculation method is simple and intuitive; the corresponding weights can be quickly obtained based on the vertex positions without complex iterative calculations, significantly improving the efficiency of weight calculation. At the same time, the granularity of sub-band division can also be flexibly adjusted according to the mesh accuracy requirements. For high-precision meshes, finer sub-bands can be divided to obtain smoother weight changes, while for scenarios with high real-time requirements, the number of sub-bands can be reduced to simplify calculations, adapting to different performance and accuracy requirements.

[0142] For example, if an intra-band interval is divided into N consecutive sub-bands, with the index of the first sub-band closest to the cutting plane being 0 and the index of the innermost sub-band being N-1, then the weight value of the vertex is (N-1-index) / (N-1). At this time, the vertex with index 0 has a weight of 1, achieving the strongest adsorption effect, while the vertex with index N-1 has a weight of 0, without producing additional deformation. This perfectly matches the strength requirement of gradient adsorption. The calculation logic is simple and efficient, making it easy to implement quickly.

[0143] In one embodiment, S146, targeting the pose-adjusted half-body reference mesh, and applying non-rigid deformation to the vertices within the seam region according to the gradient weights, may include:

[0144] S1461: Construct a local affine deformation model, define local affine transformation parameters on the mesh edges of the seam zone, and realize deformation transfer through graph propagation.

[0145] S1462: With the goal of minimizing the energy function, the local affine transformation parameters of each mesh edge in the joint zone are solved by optimizing the energy function, and the vertices in the joint zone are subjected to non-rigid deformation based on the solved local affine transformation parameters.

[0146] The energy function includes a data term, a preservation term, a stiffness term, and a Laplacian smoothing term. The data term is used to pull the vertex to the corresponding target position on the pose-adjusted half-body reference mesh, and the weight of the data term is positively correlated with the gradient weight. The preservation term is used to constrain the deformed vertex to its original position, and the weight of the preservation term is negatively correlated with the gradient weight. The stiffness term is used to constrain the local affine transformation parameters. The Laplacian smoothing term is used to preserve the local smoothness of the deformed mesh.

[0147] In this embodiment, when performing non-rigid deformation on vertices within the seam zone according to the gradient weight, this application defines local affine transformation parameters on the mesh edges to perform local deformation optimization only for the seam zone region, without needing to perform global calculations on the entire body source mesh, which greatly reduces the amount of computation and improves the efficiency of deformation calculation.

[0148] Specifically, in this application, the energy function controls the deformation result through four constraints. The data term controls the adsorption intensity by using gradient weights, making the high-weight vertices at the seams move more strongly toward the target position of the half-body reference mesh. The preservation term, on the other hand, makes the low-weight inner vertices closer to their original positions. The combination of these two terms perfectly meets the requirements of gradient adsorption. The stiffness term constrains the local affine transformation to avoid excessively strange deformations, and the Laplacian smoothing term ensures that the mesh remains smooth and continuous locally after deformation, without producing uneven or distorted meshes. This locally optimized deformation method can fully align the vertices at the seams without causing unnecessary deformation effects on other areas of the whole-body source mesh, preserving the original geometric information of the input mesh to the greatest extent, while also having low computational cost and naturally stable deformation results.

[0149] In one specific implementation, the calculation formula for the data item of the energy function in this application can be weight × gradient weight × ||deformation vertex - soft anchor point||², where the weight is a preset global weight coefficient, and the soft anchor point is the matching target position of the corresponding deformation vertex on the half-body reference mesh. This allows the gradient weight to be naturally integrated into the energy constraint, achieving the gradient adsorption effect without additional modification to the optimization framework. It has good compatibility and is easy to integrate into existing non-rigid deformation optimization processes. The calculation formula for the retention term can be anchor weight × (1 - gradient weight) × ||deformation vertex - original position||², where the anchor weight is also a preset global weight coefficient. In this way, the strength requirement of the retention term can be naturally adapted using only (1 - gradient weight), forming a complementary constraint relationship with the data item. This can be achieved without modifying the structure of the original optimization framework, adapting to existing energy-based non-rigid deformation processes.

[0150] The stiffness term in this application can be calculated using a phased decreasing stiffness weight multiplied by a local affine stiffness constraint. The phased decreasing stiffness weight means that during the optimization iteration process, the stiffness weight gradually decreases with each iteration. In the early stages, a higher stiffness weight is used to maintain the overall shape stability of the mesh, preventing significant deformation deviations from the reasonable range in the initial optimization. Later, the stiffness is gradually reduced to allow for sufficient deformation adjustment to fit the target position, balancing the stability of the deformation process with the alignment accuracy of the final result. This makes the optimization process more robust and less prone to distorted results due to local optima. The Laplace smoothing term can be calculated using a smoothing weight multiplied by a Laplace smoothing constraint. The smoothing weight is a preset global coefficient that constrains the local geometry of the mesh after deformation to maintain continuity and smoothness, avoiding local distortions or uneven vertex distribution after deformation. These constraints work together to form a complete energy optimization objective, satisfying the requirements of gradual adsorption while ensuring that the deformation result is natural and reasonable, without distortions or discontinuities.

[0151] Furthermore, the total loss in this application can be sqrt(data term + stiffness term) + Laplacian smoothing term. A dual-loop optimization of local affine parameters is used: the outer loop updates vertex positions, while the inner loop solves for local affine transformation parameters. Through hierarchical iterative optimization, stable deformation results can be obtained quickly, improving optimization efficiency while ensuring accuracy. Finally, after completing the non-rigid deformation of the vertices in the seam region using the obtained local affine transformation parameters, a fully aligned and transitioned 3D human body mesh at the seam is obtained, completing the fusion and stitching of the half-body reference mesh and the full-body source mesh.

[0152] Furthermore, the local affine deformation model in this application can be replaced with a Gaussian-based deformation model or a graph-based deformation model, which can also achieve non-rigid registration, as long as it can adapt to the strength constraints of the gradient weights and only perform local deformation for the seam area. This application does not impose specific limitations. Those skilled in the art can select a suitable non-rigid deformation model according to the computational performance and accuracy requirements, and rely on the gradient weights to control the adsorption intensity at different locations to achieve natural and smooth seam fusion.

[0153] In one embodiment, the process of determining the target location may include:

[0154] For each vertex within the seam area, take K nearest neighbor vertices on the pose-adjusted half-body reference mesh, and obtain the K nearest neighbor soft anchor point positions by distance-weighted average, and use the K nearest neighbor soft anchor point positions as the target positions; where K is a positive integer greater than or equal to 2.

[0155] In this embodiment, for each vertex within the seam area, this application can search for the K nearest vertices on the pose-adjusted half-body reference mesh. A weighted average position is calculated using the distance between vertices as the weight, serving as the matching target for that vertex. Compared to single nearest neighbor matching, this soft anchor point determination method reduces matching bias caused by errors in a single nearest neighbor search, improving the robustness of target position estimation. Even if individual nearest neighbor vertices are mismatched, the weighted average of multiple nearest neighbors can offset the error, resulting in a more stable and reasonable target position. This provides a more reliable constraint basis for subsequent deformation optimization, preventing deformation misalignment caused by incorrect matching. Furthermore, the value of K can be flexibly adjusted according to the vertex density of the half-body reference mesh. Higher vertex density allows for a larger K value to obtain a smoother target position, further improving matching stability.

[0156] In one embodiment, the mesh fusion process for the initial three-dimensional human body mesh in step S150 may include:

[0157] The upper body portion after non-rigid adsorption is merged with the lower body portion of the whole-body source mesh into a single mesh, and the single mesh is reconstructed at the seam to make the vertices and triangular faces on both sides of the seam topologically continuous.

[0158] The reconstruction process can be implemented in any one of the following ways: Poisson fusion, edge-stitched mesh fusion, or implicit surface-based mesh fusion.

[0159] In this embodiment, when performing mesh fusion on the initial 3D human body mesh, this application merges the two non-rigidly deformed parts into a single mesh. Only simple topological reconstruction at the seams is needed to achieve overall topological continuity. Different reconstruction methods can adapt to different application scenarios: Poisson fusion can quickly complete mesh reconstruction with high computational efficiency, suitable for scenarios with high real-time requirements; edge-stitched mesh fusion can accurately preserve the geometric features of the original mesh, achieving higher accuracy, suitable for high-precision human body modeling scenarios; implicit surface-based mesh fusion can better handle complex topological seams, adapting to fusion scenarios with significant morphological changes. All reconstruction methods can achieve final topological continuity, ensuring that the fused full-body 3D human body mesh has no topological breaks or gaps, meeting the topological requirements of subsequent applications.

[0160] For example, when a quick preview is needed, Poisson reconstruction can be selected to output the fused result as quickly as possible while ensuring overall consistency. When a high-precision final model is required, an edge-stitching-based fusion method can be chosen to preserve the detailed features of the original input mesh to the greatest extent. Different reconstruction methods can be flexibly selected as needed to adapt to different development and application requirements.

[0161] In one embodiment, the non-rigid adsorption of the intermediate three-dimensional human body mesh in S160 may include:

[0162] S161: In the seam zone region of the intermediate three-dimensional human body mesh, a constant adsorption weight is assigned to each vertex.

[0163] S162: Using the half-body reference mesh as a high-precision geometric target, perform non-rigid deformation on the vertices within the seam area according to the adsorption weight.

[0164] The stiffness constraint strength during the second non-rigid adsorption is less than that during the first non-rigid adsorption.

[0165] In this embodiment, since the initial non-rigid deformation and topology reconstruction have already achieved topological continuity and preliminary alignment at the seam, only secondary fine-tuning of the fused seam region is needed to better match the high-precision features of the half-body reference mesh. Therefore, only constant snapping weights need to be assigned to the vertices of the seam region, without recalculating the gradient weights, simplifying the secondary optimization process. Simultaneously, reducing the stiffness constraint strength allows for easier vertex position adjustment, fully conforming to the high-precision geometric details of the half-body reference mesh. This compensates for the insufficient adjustment of local details in the initial deformation optimization to maintain overall stability, further improving the geometric smoothness and alignment accuracy of the seam region. This results in a final output full-body 3D human mesh with overall geometric continuity and better detail consistency.

[0166] Understandably, while the seams are aligned after the initial deformation, minor local geometric deviations exist. Reducing the stiffness constraint allows vertices to be repositioned over a wider range, enabling local details to better match the surface shape of the half-body reference mesh, eliminating splicing marks, and resulting in a more natural overall mesh transition. Secondary non-rigid adsorption only optimizes the seam area locally, without needing to adjust vertex positions in other parts of the body. This results in low computational cost while significantly improving the geometric accuracy and visual effect of the final output model, offering excellent cost-effectiveness.

[0167] In one embodiment, step S160, which involves reducing the surface area and baking the texture of the adsorbed intermediate 3D human body mesh to obtain the final full-body 3D human body mesh, may include:

[0168] S163: The quadratic error metric method is used to reduce the number of triangles in the intermediate three-dimensional human body mesh after adsorption, so that the number of triangles is reduced to the target size.

[0169] S164: Automatically unwrap UVs and bake texture maps on the intermediate 3D human body mesh after surface reduction to generate a renderable 3D asset format, resulting in the final full-body 3D human body mesh.

[0170] In this embodiment, when reducing the number of faces in the intermediate three-dimensional human body mesh after adsorption, the quadratic error metric method can simplify the number of triangles while preserving the original geometric features of the mesh to the greatest extent. This avoids the problem of local feature loss or excessive surface smoothing after a large reduction in faces. It can ensure model quality while controlling the mesh size and adapt to the face count requirements of different downstream applications. For example, real-time rendering scenes can reduce the face count to a lower scale, while offline rendering scenes can retain more faces to obtain higher details.

[0171] It is understood that this application can not only use the quadratic error metric method for mesh reduction, but also other mesh simplification algorithms, such as cluster-based mesh simplification, curvature-based mesh simplification, edge-folding mesh reduction algorithms, etc. Those skilled in the art can flexibly choose according to their own requirements for the accuracy and speed of mesh simplification, as long as the number of triangular faces can be reduced to the target size while reasonably preserving the original geometric features. This application does not make any specific limitations.

[0172] After reducing the polygon count, this application can automatically unwrap UVs and bake textures to bake the color and geometric details of the original multi-part meshes into a unified texture map, generating renderable 3D assets that meet industry standards. Without the need for manual UV and texture processing, these assets can be directly used in downstream applications such as games, virtual human live streaming, and animation production, significantly improving the automation level of the entire process and reducing the manual costs of subsequent processing.

[0173] Specifically, during automatic UV unwrapping and texture baking, this application can directly bake the color information of different parts of the original mesh to generate a unified texture. Alternatively, it can bake the geometric details of the original high-precision mesh to a low-precision mesh through normal maps and height maps, allowing the low-polygon mesh to still present rich visual details, balancing rendering performance and visual effects, and adapting to different application needs. The final generated complete 3D human body mesh asset has topological continuity, natural geometric transitions, and unified texture, and can be directly used in various downstream application scenarios without additional manual adjustments.

[0174] The following describes the whole-body three-dimensional human body mesh generation device provided in the embodiments of this application. The whole-body three-dimensional human body mesh generation device described below can be referred to in correspondence with the whole-body three-dimensional human body mesh generation method described above.

[0175] In one embodiment, such as Figure 4 As shown, Figure 4 This is a schematic diagram of a full-body three-dimensional human body mesh generation device provided in an embodiment of this application; this application also provides a full-body three-dimensional human body mesh generation device, which may include an image cropping module 210, a mesh generation module 220, a mesh cropping module 230, a pose adjustment and snapping module 240, a mesh fusion and densification module 250, and a mesh reduction and baking module 260, specifically including the following:

[0176] Image cropping module 210 is used to acquire a full-body human image and crop the full-body human image to obtain an upper-body human image.

[0177] Mesh generation module 220 is used to generate a full-body source mesh corresponding to the full-body human image and a half-body reference mesh corresponding to the upper-body human image, wherein the mesh accuracy of the half-body reference mesh is higher than that of the full-body source mesh.

[0178] The mesh cropping module 230 is used to determine the effective cropping direction based on the cropping frame boundary of the upper body human image, and to crop the full-body source mesh and the half-body reference mesh based on the effective cropping direction.

[0179] The pose adjustment and adsorption module 240 is used to adjust the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh, and to non-rigidly adsorb the mesh at the seam of the cropped full-body source mesh with the pose-adjusted half-body reference mesh as the target to obtain the initial three-dimensional human body mesh.

[0180] The mesh fusion and densification module 250 is used to perform mesh fusion on the initial three-dimensional human body mesh and to densify the fused initial three-dimensional human body mesh to obtain an intermediate three-dimensional human body mesh.

[0181] The mesh reduction and baking module 260 is used to perform non-rigid adsorption on the intermediate three-dimensional human body mesh again, and to perform surface reduction and texture baking on the adsorbed intermediate three-dimensional human body mesh to obtain the final full-body three-dimensional human body mesh.

[0182] In the above embodiments, the technical route of "cropping and positioning → pose adjustment → non-rigid adsorption → fusion and densification → secondary adsorption → post-processing" is adopted. A high-precision half-body reference mesh is used as the target to perform non-rigid adsorption at the seams of the full-body source mesh, making the accuracy of the upper half of the full-body mesh comparable to that of the half-body reference mesh. In the pose adjustment step, the shell point cloud is extracted from the cropped half-body reference mesh and the full-body source mesh for rigid registration. Only the half-body reference mesh is subjected to rigid transformation, while the full-body source mesh remains stationary. Even if there is a large pose deviation between the half-body reference mesh and the full-body source mesh, the half-body reference mesh can be stably registered. Aligning the reference mesh to the whole-body coordinate system provides a good initial condition for subsequent non-rigid registration, avoiding convergence instability or local mesh folding caused by direct non-rigid registration. Next, this application ensures the overall topological continuity of the whole-body mesh through mesh fusion, eliminating seam traces. Then, a second non-rigid registration is performed using the half-body reference mesh as a high-precision geometric target, effectively restoring the local high-frequency details lost during the fusion process and compensating for the accuracy loss of a single fusion. Finally, after face reduction and texture baking, a renderable 3D asset with a moderate number of triangles and UV texture mapping is output, meeting the direct use requirements of downstream applications.

[0183] In one embodiment, the mesh trimming module 230 includes:

[0184] The effective pixel value determination module is used to determine whether there are effective pixel values ​​at the top, bottom, left, and right boundaries of the cropping frame of the upper body human image.

[0185] The effective cropping direction determination module is used to determine the boundary with valid pixel values ​​as the effective boundary, and to map the effective boundary as the effective cropping direction in three-dimensional space according to the imaging perspective parameters of the whole body image.

[0186] In one embodiment, the effective clipping direction is configured using a clipping surface data structure, with each effective clipping direction corresponding to a clipping surface.

[0187] The cutting surface includes four parameters: arrangement axis information, scanning axis, scanning direction, and band length ratio.

[0188] The arrangement axis information is composed of the arrangement order of the other two coordinate axes besides the scanning axis among the three coordinate axes x, y, and z. The scanning axis is the x-axis, y-axis, or z-axis. The scanning direction is the positive or negative direction along the scanning axis. The band length ratio is the proportion of the clipping surface along the scanning axis direction to the span of the target mesh bounding box, used to determine the end position of the clipping surface along the scanning axis direction.

[0189] In one embodiment, the mesh trimming module 230 includes:

[0190] The out-of-band far-end region determination module is used to define the set of vertices that extend beyond the endpoint of the clipping surface along the corresponding scan direction as the out-of-band far-end region of the clipping surface corresponding to the effective clipping direction; wherein, when there are multiple effective clipping directions, the out-of-band far-end region is the intersection of the out-of-band far-end regions of the clipping surfaces corresponding to each effective clipping direction.

[0191] The full-body source mesh trimming module is used to trim the full-body source mesh by deleting vertices and associated triangular faces belonging to the outer-band far-end region of the full-body source mesh, while retaining the area near the seams and the main body.

[0192] The half-body reference mesh clipping module is used to delete vertices in the half-body reference mesh that do not belong to the out-of-band far-end region when clipping the half-body reference mesh, and only retain the mesh portion of the half-body reference mesh that belongs to the out-of-band far-end region.

[0193] In one embodiment, the pose adjustment and adsorption module 240 includes:

[0194] The vertex extraction module is used to extract vertices falling within the outward expansion range of the axis-aligned bounding box from the cropped half-body reference mesh as a first set of vertices, and extract vertices falling outside the inward contraction range of the axis-aligned bounding box from the cropped full-body source mesh as a second set of vertices.

[0195] The vertex count determination module is used to determine whether the number of vertices in the second point set has reached a preset threshold.

[0196] The pose adjustment module is used to estimate the rigid transformation matrix between the first point set and the second point set through a rigid registration algorithm if the condition is met, and apply the rigid transformation matrix to the half-body reference mesh to adjust the pose of the half-body reference mesh.

[0197] The pose-not-adjustment module is used to otherwise keep the half-body reference mesh in its original pose.

[0198] In one embodiment, the pose adjustment and adsorption module 240 includes:

[0199] The gradient weight calculation module is used to calculate the gradient weight of each vertex in the seam zone area of ​​the cropped full-body source mesh. The gradient weight decreases monotonically from the seam towards the inside of the seam zone.

[0200] The non-rigid deformation module is used to perform non-rigid deformation on the vertices in the seam area based on the gradient weight, with the pose-adjusted half-body reference mesh as the target, so that the adsorption strength of the vertices near the seam is greater than that of the vertices far from the seam.

[0201] In one embodiment, the gradient weight calculation module includes:

[0202] The cutting surface determination module is used to determine the cutting surface corresponding to the effective cutting direction.

[0203] A vertex determination module is used to determine, based on the cut surface, vertices that meet qualification criteria within the seam zone region of the cut full-body source mesh. The qualification criteria include: the vertex is not excluded from the far-end region outside the band by any of the cut surfaces, and falls at least within the band region of any of the cut surfaces.

[0204] The first weight calculation module is used to calculate the weight value of each vertex in each clipping plane when the vertex falls within the in-band interval of multiple clipping planes, based on the vertex's position in the in-band interval of each clipping plane, and take the minimum value among the weight values ​​as the gradient weight of the vertex.

[0205] The second weight calculation module is used to calculate the weight value of the vertex in the clipping plane when the vertex falls only within the band of a clipping plane, and to use the calculated weight value as the gradient weight of the vertex.

[0206] In one embodiment, the second weight calculation module includes:

[0207] The sub-band index determination module is used to divide the intra-band interval of the clipping surface into several sub-bands along the scanning direction and determine the sub-band index to which the vertex belongs in the intra-band interval.

[0208] The weight value calculation module is used to normalize the relative position of the sub-band index in the band interval to obtain the weight value of the vertex in the cutting surface. The weight value decreases from the outside closer to the seam surface to the inside farther away from the seam surface.

[0209] In one embodiment, the non-rigid deformation module includes:

[0210] The local affine transformation parameter definition module is used to construct a local affine deformation model. Local affine transformation parameters are defined on the mesh edges of the joint zone, and deformation is transferred through graph propagation.

[0211] The vertex deformation module is used to minimize the energy function by optimizing the energy function to solve the local affine transformation parameters of each mesh edge in the seam zone, and to perform non-rigid deformation on the vertices in the seam zone according to the solved local affine transformation parameters.

[0212] The energy function includes a data term, a preservation term, a stiffness term, and a Laplacian smoothing term. The data term is used to pull the vertex to the corresponding target position on the pose-adjusted half-body reference mesh, and the weight of the data term is positively correlated with the gradient weight. The preservation term is used to constrain the deformed vertex to its original position, and the weight of the preservation term is negatively correlated with the gradient weight. The stiffness term is used to constrain the local affine transformation parameters. The Laplacian smoothing term is used to preserve the local smoothness of the deformed mesh.

[0213] In one embodiment, the process of determining the target location includes:

[0214] The target location determination module is used to take K nearest neighbor vertices on the pose-adjusted half-body reference mesh for each vertex in the seam area, obtain the K nearest neighbor soft anchor point positions by distance-weighted average, and use the K nearest neighbor soft anchor point positions as the target location.

[0215] Where K is a positive integer greater than or equal to 2.

[0216] In one embodiment, the mesh fusion and densification module 250 performs mesh fusion on the initial three-dimensional human body mesh, including:

[0217] The mesh fusion module is used to merge the upper body part after non-rigid adsorption with the lower body part of the whole body source mesh into a single mesh, and to reconstruct the single mesh at the seam so that the vertices and triangular faces on both sides of the seam are topologically continuous.

[0218] The reconstruction process can be implemented in any one of the following ways: Poisson fusion, edge-stitched mesh fusion, or implicit surface-based mesh fusion.

[0219] In one embodiment, the mesh reduction and baking module 260 performs non-rigid adsorption on the intermediate three-dimensional human body mesh again, including:

[0220] The adsorption weight allocation module is used to assign a constant adsorption weight to each vertex within the seam zone of the intermediate three-dimensional human body mesh.

[0221] The secondary deformation module is used to perform non-rigid deformation on the vertices in the seam area according to the adsorption weight, using the half-body reference mesh as a high-precision geometric target.

[0222] The stiffness constraint strength during the second non-rigid adsorption is less than that during the first non-rigid adsorption.

[0223] In one embodiment, the mesh reduction and baking module 260 performs surface reduction and texture baking on the adsorbed intermediate 3D human body mesh to obtain the final full-body 3D human body mesh, including:

[0224] The mesh reduction module is used to reduce the number of triangles in the middle 3D human body mesh after adsorption by using an edge folding simplification algorithm, thereby reducing the number of triangles to the target size.

[0225] The mesh baking module is used to automatically unwrap the UVs and bake texture maps on the intermediate 3D human body mesh after the polygon reduction, generating a renderable 3D asset format to obtain the final full-body 3D human body mesh.

[0226] In one embodiment, this application also provides a computer-readable storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the full-body three-dimensional human mesh generation method as described in any of the above embodiments.

[0227] In one embodiment, this application also provides a computer device, including: one or more processors, and memory.

[0228] The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the full-body three-dimensional human mesh generation method as described in any of the above embodiments.

[0229] Indicatively, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 5 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions executable by the processing component 302, such as application programs. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the full-body 3D human mesh generation method of any of the above embodiments.

[0230] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0231] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0232] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0233] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0234] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating a full-body three-dimensional human body mesh, characterized in that, The method includes: A full-body image is acquired, and the full-body image is cropped to obtain an upper-body image. Generate a full-body source mesh corresponding to the full-body human image and a half-body reference mesh corresponding to the upper-body human image, wherein the half-body reference mesh has a higher mesh precision than the full-body source mesh; The effective cropping direction is determined based on the cropping box boundary of the upper body human image, and the full-body source mesh and the half-body reference mesh are cropped based on the effective cropping direction; Based on the cropped full-body source mesh and the cropped half-body reference mesh, the pose of the half-body reference mesh is adjusted, and the mesh at the seam of the cropped full-body source mesh is non-rigidly snapped with the pose-adjusted half-body reference mesh as the target to obtain the initial three-dimensional human body mesh. The initial three-dimensional human body mesh is fused, and the fused initial three-dimensional human body mesh is densified to obtain an intermediate three-dimensional human body mesh. The intermediate three-dimensional human body mesh is then subjected to non-rigid adsorption again, and the adsorbed intermediate three-dimensional human body mesh is subjected to surface reduction and texture baking to obtain the final full-body three-dimensional human body mesh.

2. The method according to claim 1, characterized in that, The step of determining the effective cropping direction based on the cropping frame boundary of the upper body human image includes: Determine whether there are valid pixel values ​​at the top, bottom, left, and right boundaries of the cropping frame of the upper body image; The boundary with valid pixel values ​​is determined as the valid boundary, and the valid boundary is mapped to the effective cropping direction in three-dimensional space according to the imaging viewpoint parameters of the whole body image.

3. The method according to claim 1, characterized in that, The effective clipping direction is configured using a clipping surface data structure, and each effective clipping direction corresponds to a clipping surface. The cutting surface includes four parameters: arrangement axis information, scanning axis, scanning direction, and band length ratio. The arrangement axis information is composed of the arrangement order of the other two coordinate axes besides the scanning axis among the three coordinate axes x, y, and z. The scanning axis is the x-axis, y-axis, or z-axis. The scanning direction is the positive or negative direction along the scanning axis. The band length ratio is the proportion of the clipping surface along the scanning axis direction to the span of the target mesh bounding box, used to determine the end position of the clipping surface along the scanning axis direction.

4. The method according to claim 3, characterized in that, The step of cropping the full-body source mesh and the half-body reference mesh based on the effective cropping direction includes: For the clipping surface corresponding to the effective clipping direction, the set of vertices that extend beyond the endpoint of the clipping surface along the corresponding scan direction coordinates is taken as the out-of-band far-end region of the clipping surface; wherein, when there are multiple effective clipping directions, the out-of-band far-end region is the intersection of the out-of-band far-end regions of the clipping surfaces corresponding to each effective clipping direction. When trimming the full-body source mesh, delete the vertices and associated triangles belonging to the out-of-band far-end region in the full-body source mesh, and retain the area near the seam and the main body; When clipping the half-body reference mesh, delete the vertices in the half-body reference mesh that do not belong to the out-of-band far-end region, and only retain the mesh portion in the half-body reference mesh that belongs to the out-of-band far-end region.

5. The method according to claim 1, characterized in that, The step of adjusting the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh includes: Based on the axis-aligned bounding box of the cropped half-body reference mesh, extract the vertices that fall within the outward expansion range of the axis-aligned bounding box from the cropped full-body source mesh as the first point set, and extract the vertices that fall outside the inward contraction range of the axis-aligned bounding box from the cropped half-body reference mesh as the second point set. Determine whether the number of vertices in the second point set has reached a preset threshold. If so, the rigid transformation matrix between the first point set and the second point set is estimated by the rigid registration algorithm, and the rigid transformation matrix is ​​applied to the half-body reference mesh to adjust the pose of the half-body reference mesh; Otherwise, the half-body reference mesh is kept in its original pose.

6. The method according to claim 1, characterized in that, The step of using the pose-adjusted half-body reference mesh as the target and performing non-rigid snapping on the seams of the cropped full-body source mesh includes: Within the seam zone of the cropped full-body source mesh, the gradient weight of each vertex is calculated, and the gradient weight decreases monotonically from the seam towards the inside of the seam zone. Using the pose-adjusted half-body reference mesh as the target, the vertices in the seam area are subjected to non-rigid deformation according to the gradient weight, so that the adsorption strength of the vertices near the seam is greater than that of the vertices far from the seam.

7. The method according to claim 6, characterized in that, The step of calculating the gradient weights of each vertex within the seam zone of the cropped full-body source mesh includes: Determine the cutting surface corresponding to the effective cutting direction; Within the seam zone of the trimmed full-body source mesh, vertices that meet the eligibility criteria are determined based on the trimming surfaces. The eligibility criteria include: the vertex is not excluded from the far-end region outside the strip by any of the trimming surfaces, and falls at least within the strip zone of any of the trimming surfaces. For each vertex, when the vertex falls within the in-band interval of multiple clipping planes, the weight value of the vertex in each clipping plane is calculated based on the position of the vertex in the in-band interval of each clipping plane, and the minimum value among the weight values ​​is taken as the gradient weight of the vertex. When a vertex falls within the band of only one clipping plane, calculate the weight value of the vertex in that clipping plane and use the calculated weight value as the gradient weight of the vertex.

8. The method according to claim 7, characterized in that, The calculation of the weight value of the vertex in the clipping plane includes: The intra-band interval of the clipping surface is divided into several sub-bands along the scanning direction, and the sub-band index to which the vertex belongs in the intra-band interval is determined; The weight value of the vertex in the cutting plane is obtained by normalizing the relative position of the sub-band index in the interval within the band, and the weight value decreases from the outside closer to the seam surface to the inside farther away from the seam surface.

9. The method according to claim 6, characterized in that, The step of using the pose-adjusted half-body reference mesh as the target and applying non-rigid deformation to the vertices within the seam area according to the gradient weights includes: A local affine deformation model is constructed, and local affine transformation parameters are defined on the mesh edges of the seam zone. Deformation is transferred through graph propagation. With the goal of minimizing the energy function, the local affine transformation parameters of each mesh edge in the joint zone are solved by optimizing the energy function, and the vertices in the joint zone are subjected to non-rigid deformation based on the solved local affine transformation parameters. The energy function includes a data term, a preservation term, a stiffness term, and a Laplacian smoothing term. The data term is used to pull the vertex to the corresponding target position on the pose-adjusted half-body reference mesh, and the weight of the data term is positively correlated with the gradient weight. The preservation term is used to constrain the deformed vertex to its original position, and the weight of the preservation term is negatively correlated with the gradient weight. The stiffness term is used to constrain the local affine transformation parameters. The Laplacian smoothing term is used to preserve the local smoothness of the deformed mesh.

10. The method according to claim 9, characterized in that, The process of determining the target location includes: For each vertex in the seam area, take K nearest neighbor vertices on the pose-adjusted half-body reference mesh, and obtain the K nearest neighbor soft anchor point positions by distance-weighted average, and use the K nearest neighbor soft anchor point positions as the target positions; Where K is a positive integer greater than or equal to 2.

11. The method according to claim 1, characterized in that, The mesh fusion of the initial 3D human body mesh includes: The upper body part after non-rigid adsorption is merged with the lower body part of the whole body source mesh into a single mesh, and the single mesh is reconstructed at the seam to make the vertices and triangular faces on both sides of the seam topologically continuous. The reconstruction process can be implemented in any one of the following ways: Poisson fusion, edge-stitched mesh fusion, or implicit surface-based mesh fusion.

12. The method according to claim 1, characterized in that, The process of applying non-rigid adsorption to the intermediate three-dimensional human body mesh again includes: Within the seam zone of the intermediate three-dimensional human body mesh, a constant adsorption weight is assigned to each vertex. Using the half-body reference mesh as a high-precision geometric target, non-rigid deformation is applied to the vertices within the seam area according to the adsorption weight; The stiffness constraint strength during the second non-rigid adsorption is less than that during the first non-rigid adsorption.

13. The method according to claim 1, characterized in that, The process of reducing the surface area and baking the texture of the adsorbed intermediate 3D human body mesh to obtain the final full-body 3D human body mesh includes: The quadratic error metric method is used to reduce the number of triangular faces in the intermediate three-dimensional human body mesh after adsorption, thereby reducing the number of triangular faces to the target size. The intermediate 3D human body mesh after surface reduction is automatically unwrapped and texture maps are baked to generate a renderable 3D asset format, resulting in the final full-body 3D human body mesh.

14. A full-body three-dimensional human mesh generation device, characterized in that, include: The image cropping module is used to acquire a full-body human image and crop the full-body human image to obtain an upper-body human image. The mesh generation module is used to generate a full-body source mesh corresponding to the full-body human image and a half-body reference mesh corresponding to the upper-body human image, wherein the mesh accuracy of the half-body reference mesh is higher than that of the full-body source mesh. The mesh cropping module is used to determine the effective cropping direction based on the cropping box boundary of the upper body human image, and to crop the full-body source mesh and the half-body reference mesh based on the effective cropping direction. The pose adjustment and adsorption module is used to adjust the pose of the half-body reference mesh based on the cropped full-body source mesh and the cropped half-body reference mesh, and to non-rigidly adsorb the mesh at the seam of the cropped full-body source mesh with the pose-adjusted half-body reference mesh as the target to obtain the initial three-dimensional human body mesh. The mesh fusion and densification module is used to perform mesh fusion on the initial three-dimensional human body mesh and to densify the fused initial three-dimensional human body mesh to obtain an intermediate three-dimensional human body mesh. The mesh reduction and baking module is used to perform non-rigid adsorption on the intermediate three-dimensional human body mesh again, and to perform surface reduction and texture baking on the adsorbed intermediate three-dimensional human body mesh to obtain the final full-body three-dimensional human body mesh.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the whole-body three-dimensional human mesh generation method as described in any one of claims 1 to 13.

16. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the full-body three-dimensional human mesh generation method as described in any one of claims 1 to 13.

Images