Semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling

Abstract

The invention provides a semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling, and the method comprises the following steps: initializing 3D Gaussian model parameters through a multi-view image, and dividing the parameters into an identity group, a geometric group and an appearance group according to semantics; performing selective freezing or updating on each group of parameters in different training stages by adopting a staged training strategy and combining a differential updating mechanism; a local calibration mechanism is introduced in the geometric optimization stage, a bone correction strategy is activated through triggering conditions such as key part offset evaluation, symmetry detection and attitude anomaly analysis, and precise adjustment of a local structure is achieved; according to the method, the technical problems of serious parameter coupling, identity feature distortion and high data dependence in small sample training in the prior art are effectively solved, and the robustness and generalization ability of the model under limited input data are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of video transmission technology, and in particular to a semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling. Background Art

[0002] Digital humans refer to virtual characters with human appearance or behavior produced by computer technology and artificial intelligence technology. These virtual characters have a wide range of application scenarios and have the advantages of high realism, strong flexibility and high cost-effectiveness. 3D Gaussian Sputtering (3DGS) is an explicit three-dimensional modeling and real-time rendering method that has developed rapidly in recent years. This method represents the scene as a set of point clouds with Gaussian distribution attributes. Each point contains parameters such as position p, scaling matrix S, rotation matrix R, transparency α, color information (such as spherical harmonic coefficients SH coefficients). Through an efficient rasterization rendering process, 3DGS can achieve high-quality and real-time three-dimensional reconstruction effects, showing great potential in applications such as digital human modeling, virtual anchors, and remote conferencing. Compared with the neural radiance field (NeRF) method based on implicit representation, 3DGS has significant advantages in rendering speed, usually reaching more than 30FPS, meeting the performance requirements of interactive systems. At the same time, its explicit point cloud structure is also easier to edit, drive and dynamically control.

[0003] Although 3DGS achieves a good balance between efficiency and quality, it still faces several key technical bottlenecks in practical applications, especially in small sample training scenarios:

[0004] (1) Parameter coupling problem: The 3DGS model suffers from an imbalance in the gradient distribution of the identity group (p, S), geometry group (R, α), and appearance group (SH) (68.2% vs 5%), which leads to limited dynamic expression and attribute interference;

[0005] (2) Identity characteristics are easily damaged: Traditional full-parameter optimization is prone to identity feature deviation under small sample sizes, and the FID index fluctuates significantly.

[0006] Therefore, the current relevant technologies and application research on the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling still need to be further improved. Summary of the Invention

[0007] In view of this, the present invention proposes a semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling, which solves the technical problems in the existing technology such as severe parameter coupling, easy distortion of identity features, and strong data dependence.

[0008] The technical solution of the present invention is achieved as follows:

[0009] First, a semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling includes the following steps:

[0010] Initialize the 3D Gaussian model parameters through multi-view images, and divide the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics;

[0011] Training is done in stages, using a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group;

[0012] Based on the training results, a 3D digital human model is generated.

[0013] Based on this technical solution, the 3D Gaussian model parameters are divided into identity group, geometry group and appearance group according to semantics, among which,

[0014] The identity group includes position parameters p and scaling matrix S, which are used to maintain the basic identity features of the digital human;

[0015] The geometry group includes a rotation matrix R and a transparency parameter α, which are used to control the posture and spatial structure;

[0016] The appearance group includes spherical harmonic coefficients SH, which are used to manage surface lighting and texture information.

[0017] Based on this technical solution, the phases are divided into a geometric optimization phase and an appearance optimization phase, specifically including:

[0018] Geometry optimization stage: freeze the 3D Gaussian model parameters of the identity group and appearance group, and update the 3D Gaussian model parameters of the geometry group;

[0019] Appearance optimization stage: freeze the 3D Gaussian model parameters of the identity group and geometry group, and update the 3D Gaussian model parameters of the appearance group.

[0020] Based on this technical solution, a local calibration mechanism is introduced in the geometric optimization stage. The triggering conditions of the local calibration mechanism include:

[0021] Key part offset: The Euclidean distance between the key points of the face or hand and the standard template is greater than the set threshold;

[0022] Symmetry detection failed: the spatial position difference of the symmetrical structure exceeds the preset threshold;

[0023] Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds the preset threshold;

[0024] The local calibration process is activated when one of the trigger conditions, namely, key part deviation, symmetry detection failure, or posture abnormality detection, is met.

[0025] Based on this technical solution, the skeleton correction strategy, when the trigger conditions are met, optimizes the mapping function and weights through supervised learning, uses the skeleton correction formula to fine-tune the Gaussian point position in the local area, and resets the geometric group learning rate. The calibration formula is:

[0026]

[0027] Among them, P is the original position of the Gaussian point, P corrected is the corrected position, N is the number of joints involved in the transformation, w k is the skin weight of the kth joint, T k (ΔR) is the displacement mapping function corresponding to the kth joint.

[0028] The displacement mapping function is:

[0029] T k (ΔR)=W k ·log(ΔR k )+b k

[0030] Among them, ΔR represents the rotation change between the target posture and the current posture, log(ΔR k ) is the axis-angle representation of the rotation matrix, W k is a learnable linear transformation matrix, b k is the base offset.

[0031] In a second aspect, the present invention provides a semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling, which applies the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling as described in any one of the first aspects, comprising:

[0032] The grouping module initializes the 3D Gaussian model parameters through multi-view images and divides the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics;

[0033] The training module is used for staged training and uses a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group;

[0034] The output module is used to generate a 3D digital human model based on the training results.

[0035] On the basis of this technical solution, the training module also includes:

[0036] A differential update unit, used to update each set of parameters in stages;

[0037] A freeze control unit, which manages full freezes of identity groups and staged freezes of other groups;

[0038] Local calibration unit: used to detect model structure deviations and trigger skeleton correction strategies;

[0039] Skeleton correction unit: used to perform geometric parameter correction of the formula.

[0040] Based on this technical solution, the grouping module includes identity group, geometry group and appearance group, wherein,

[0041] The identity group is used to maintain the basic identity characteristics of the digital human by fixing the center position and scaling matrix;

[0042] Geometry group: used to control the pose and spatial structure by optimizing the rotation matrix and transparency;

[0043] Appearance Group: Used to manage surface lighting and textures by updating spherical harmonic coefficients.

[0044] Based on this technical solution, the local calibration unit is used to detect the following structural anomalies during the geometry optimization phase and serve as a basis for triggering the skeletal correction strategy:

[0045] Key part offset: The Euclidean distance between the key points of facial features or hands and the standard template is greater than the set threshold;

[0046] Symmetry detection failed: the spatial position difference of the symmetrical structure exceeds the preset threshold;

[0047] Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds the preset threshold;

[0048] When one of the trigger conditions of key part offset, symmetry detection failure or posture abnormality detection is met, the skeleton correction unit is activated to perform the geometric parameter correction operation based on the joint transformation function.

[0049] In a third aspect, the present invention provides a computer-storable medium storing instructions, which, when executed on a computer, executes the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling as described in any one of the first aspects.

[0050] The semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling described in the present invention have the following advantages over the prior art:

[0051] Parameter semantic grouping: The parameters of the 3D Gaussian sputtering model are divided into three independent groups according to their physical meaning and function: identity group, geometry group, and appearance group. By introducing prior knowledge of human anatomy for semantic division, each group of parameters has stronger interpretability and controllability during the optimization process, avoiding mutual interference between parameters.

[0052] Parameter freezing and differential update mechanism: Through the semantic grouping differential update mechanism, the identity / appearance group is frozen in the geometry stage and only the geometry group is optimized. The identity / geometry group is frozen in the appearance stage and only the SH is optimized. The parameter optimization is decoupled in stages, which improves the stability and generalization ability of the model under limited data. By freezing the identity group parameters throughout the process and only making limited adjustments to the position under skeleton drive, the identity fidelity can be greatly improved and the consistency of the digital human image can be guaranteed.

[0053] Local calibration mechanism improves modeling accuracy: To address structural deviations introduced by the initial point cloud, this paper introduces a trigger mechanism based on structural deviation detection (such as offset of key parts and symmetry failure). By activating the skeleton correction unit and fine-tuning the geometric parameters based on the joint transformation function, accurate local structure repair is achieved. This mechanism effectively addresses common issues such as identity distortion and structural asymmetry in traditional methods, further improving the quality and credibility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 This is a flow chart of the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling described in the present invention;

[0056] Figure 2 This is an architectural diagram of the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling described in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0058] The present invention aims to address the limitations of existing 3D digital human modeling technology in small-sample training scenarios. Current mainstream methods such as Neural Radiance Field (NeRF) and explicit point cloud modeling (e.g., PointΑvαtαr) face significant challenges when processing small amounts of input data, including unbalanced parameter updates, easy distortion of identity features, and a high dependence on large amounts of training data. To overcome these difficulties, the present invention proposes an innovative semantically grouped 3D Gaussian sputtering small-sample training method and system for digital human modeling. This method is particularly suitable for quickly constructing high-quality 3D digital human models with stable identity features from a small number of images.

[0059] Example 1

[0060] Semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling, such as Figure 1 As shown, the following steps are included:

[0061] Initialize the 3D Gaussian model parameters through multi-view images, and divide the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics;

[0062] Training is done in stages, using a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group;

[0063] Based on the training results, a 3D digital human model is generated.

[0064] Specifically, the 3D Gaussian model parameters are divided into identity group, geometry group and appearance group according to semantics, where

[0065] The identity group includes position parameters p and scaling matrix S, which are used to maintain the basic identity features of the digital human;

[0066] The geometry group includes a rotation matrix R and a transparency parameter α, which are used to control the posture and spatial structure;

[0067] The appearance group includes spherical harmonic coefficients SH, which are used to manage surface lighting and texture information.

[0068] Specifically, the stages include the geometry optimization stage and the appearance optimization stage, which include:

[0069] Geometry optimization stage: freeze the 3D Gaussian model parameters of the identity group and appearance group, update the 3D Gaussian model parameters of the geometry group, use the geometric loss function to prioritize restoring the structural consistency of the model, and accelerate the convergence process;

[0070] Appearance optimization stage: freeze the 3D Gaussian model parameters of the identity group and geometry group, update the 3D Gaussian model parameters of the appearance group, and use the complete rendering loss function to optimize surface lighting and color details to improve visual performance.

[0071] Specifically, the geometric optimization stage also includes a skeleton correction strategy for correcting the structural deviation caused by the initial point cloud. The triggering conditions are:

[0072] Key part offset: The Euclidean distance between the key points of facial features or hands and the standard template is greater than the set threshold (for example, >2mm);

[0073] Symmetry detection failed: The spatial position difference of symmetrical structures such as the left and right eyes and mouth corners exceeds the preset threshold (e.g. >3mm);

[0074] Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds the preset threshold (such as yaw angle > 15°);

[0075] Trigger logic:

[0076] When one of the above conditions is met, the local calibration process is activated and the bone correction strategy is called to fine-tune the position of the Gaussian points in the corresponding area.

[0077] Specifically, when the trigger conditions are met, the skeleton correction strategy optimizes the mapping function and weights through supervised learning, uses the skeleton correction formula to fine-tune the Gaussian point position in the local area, and resets the geometric group learning rate. The calibration formula is:

[0078]

[0079] Among them, P is the original position of the Gaussian point, P corrected is the corrected position, N is the number of joints involved in the transformation, w k is the skin weight of the kth joint, T k (ΔR) is the displacement mapping function corresponding to the kth joint.

[0080] The displacement mapping function is:

[0081] T k (ΔR)=W k ·log(ΔR k )+b k

[0082] Among them, ΔR represents the rotation change between the target posture and the current posture, log(ΔR k ) is the axis-angle representation of the rotation matrix, W k is a learnable linear transformation matrix, b k is the base offset.

[0083] Example 2

[0084] A semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling, comprising:

[0085] The grouping module initializes the 3D Gaussian model parameters through multi-view images and divides the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics;

[0086] The training module is used for staged training and uses a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group;

[0087] The output module is used to generate a 3D digital human model based on the training results.

[0088] Specifically, the training module also includes:

[0089] A differential update unit, used to update each set of parameters in stages;

[0090] A freeze control unit, which manages full freezes of identity groups and staged freezes of other groups;

[0091] Local calibration unit: used to detect model structure deviations and trigger skeleton correction strategies;

[0092] Skeleton correction unit: used to perform geometric parameter correction of the formula.

[0093] Based on this technical solution, the grouping module includes identity group, geometry group and appearance group, wherein,

[0094] The identity group is used to maintain the basic identity characteristics of the digital human by fixing the center position and scaling matrix;

[0095] Geometry group: used to control the pose and spatial structure by optimizing the rotation matrix and transparency;

[0096] Appearance Group: Used to manage surface lighting and textures by updating spherical harmonic coefficients.

[0097] Based on this technical solution, the local calibration unit is used to detect the following structural anomalies during the geometry optimization phase, including:

[0098] Key part offset: The Euclidean distance between the key points of facial features or hands and the standard template is greater than the set threshold (for example, >2mm);

[0099] Symmetry detection failed: The spatial position difference of symmetrical structures such as the left and right eyes and mouth corners exceeds the preset threshold (e.g. >3mm);

[0100] Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds a preset threshold (e.g., yaw angle > 15°);

[0101] In a preferred embodiment, Figure 2 As shown in FIG, the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling includes the following steps:

[0102] Initialize the 3D Gaussian model and extract point cloud data from multi-view images using SfM or MVS technology;

[0103] Initialize the Gaussian point parameters: position P, determined by the point cloud coordinates; scaling matrix S, initialized using isotropy; rotation matrix R, based on neighborhood normal vector fitting; transparency α, assessed by visibility based on point density; spherical harmonic coefficients SH, estimated by ambient lighting. Group these parameters by semantics;

[0104] According to different groups, the differential update mechanism is used to freeze or update the parameters of each group.

[0105] The training process is divided into stages as follows:

[0106] Phase 1: Geometry optimization phase.

[0107] Prioritize geometric structure recovery: Freeze the identity and appearance group parameters, only update the geometry group parameters (R, α), and use a geometric loss function. Prioritize restoring the model's geometric structure (depth and normal vectors) to avoid appearance interference and accelerate geometric convergence.

[0108] L geo =λ1L depth +λ2L normal (λ1+λ2=1)

[0109] Where depth is the depth map, which records the distance from each pixel to the camera. normal is the normal vector of a point on the surface, indicating the surface orientation of the point. depth and L normal The loss can be calculated using mean squared error (MSE) or mean absolute error (MAE).

[0110] Perform local model calibration: Activate the local calibration process based on structural deviation detection strategies, such as key part offset, symmetry detection failure, and posture anomaly detection. Specific trigger conditions include:

[0111] Key part offset (Euclidean distance between key points and standard template>5mm);

[0112] Symmetry detection failed (the spatial position difference of symmetrical structures is >3mm);

[0113] Posture anomaly detection (yaw angle > 15°).

[0114] Skeleton correction strategy: When the trigger conditions are met, the mapping function and weights are optimized through supervised learning, the position of the Gaussian points in the local area is fine-tuned using the skeleton correction formula, and the geometry group learning rate is reset to 0.001.

[0115] The core calibration formula is:

[0116]

[0117] Among them, P is the original position of the Gaussian point, P corrected is the corrected position, N is the number of joints involved in the transformation, w k is the skin weight of the kth joint, T k (ΔR) is the displacement mapping function corresponding to the kth joint.

[0118] The displacement mapping function is:

[0119] T k (ΔR)=W k ·log(ΔR k )+b k

[0120] Among them, ΔR represents the rotation change between the target posture and the current posture, log(ΔR k ) is the axis-angle representation of the rotation matrix, W k is a learnable linear transformation matrix, b k is the base offset.

[0121] Phase 2: Appearance optimization phase.

[0122] Freeze the identity group (p, S) and unfreeze the appearance group (SH). Update only the spherical harmonic coefficients, using the full loss function. Optimize surface color and lighting details, avoiding coupling between geometry and appearance properties (e.g., lighting changes affecting geometry) through staged optimization.

[0123]

[0124] Among them, rgb rendered is the RGB color value predicted by the model, rgb gt It is the RGB color value of the real scene.

[0125] Adaptive density control is performed, and the deletion condition of Gaussian sphere is transparency α<0.01.

[0126] The splitting conditions are:

[0127]

[0128] Output the optimized model to generate a 3D digital human model that can be rendered in real time.

[0129] The above method can solve the problem of parameter coupling leading to optimization failure in the existing technology. Specifically, the 3DGS model has high parameter dimensionality, usually including the identity group (center position p and scaling matrix S), the geometry group (rotation matrix R and transparency α), and the appearance group (0-3 order spherical harmonic coefficients SH). Experimental measurements show that the gradient distribution between different semantic groups during the fine-tuning process is extremely unbalanced. The geometry group parameters (R, α) dominate the geometry optimization stage, accounting for 68.2% of the gradient; the identity group parameters (p, S) are second (about 25%); and the appearance group parameters (SH) are updated less, accounting for only about 5%. This unbalanced gradient distribution leads to limited dynamic expression (for example, expression-driven lip movement leads to blurred color and texture details due to insufficient SH parameter updates) and attribute coupling interference (the strong gradients of the identity group parameters p, S may mask the optimization needs of the geometry group parameters R, α). The above problems can be alleviated by introducing a differential update mechanism based on semantic grouping. In the staged training strategy, the identity group (p, S) and appearance group (parameters are frozen in the geometry optimization stage, and only the geometry group is updated and the geometry group parameters are adjusted in combination with the model local calibration strategy (key part offset, symmetry detection failure, posture abnormality detection trigger), giving priority to restoring the model structure and avoiding attribute coupling; in the appearance optimization stage, the identity group (p, S) and geometry group (R, α) parameters are frozen, and only the appearance group (SH) is unfrozen and the complete loss function L is used. full Optimizing surface color and lighting details. This strategy ensures independent optimization of parameters for each semantic group through staged freezing and thawing, thereby alleviating the problem of gradient distribution imbalance and improving the model's performance in dynamic expressions (such as expression and lighting);

[0130] Furthermore, identity features are easily destroyed; traditional 3DGS training methods mostly adopt a full-parameter joint optimization strategy, which is easy to cause identity feature offset under the condition of a small amount of input data. When measuring identity consistency with the FID indicator, it is found that the FID value change before and after fine-tuning can reach 80%, indicating that the model has seriously interfered with the original identity features during the optimization process. The differential update mechanism based on semantic grouping freezes the identity group parameters (p, S) in both stages and only uses the skeleton correction formula in the geometric optimization stage. Limited adjustments to position P can significantly reduce the risk of identity feature drift, thereby improving the identity fidelity of the digital human.

[0131] Table 1 Efficiency optimization of Example 1

[0132]

[0133] Note: All tests were performed on the NVIDIA RTX 4090 GPU + FaceScape dataset environment.

[0134] It can be seen that the method described in the present invention has significantly improved the efficiency of training data, inference speed and video memory occupancy.

[0135] In summary, the present invention provides a semantic grouping 3D Gaussian sputtering small sample training method and system for digital human modeling. In a small sample training scenario, only a small amount of input data is required to quickly build a high-quality three-dimensional digital human model, which greatly reduces the dependence on large-scale training data. Secondly, through parameter grouping and freezing strategies, the loss of identity features is effectively prevented and the identity consistency of the model is improved. Finally, combined with the local skeleton-driven correction mechanism, the key parts are fine-tuned to further improve the accuracy and symmetry of the model in terms of geometric structure. The present invention is not only suitable for real-time interactive applications such as virtual anchors, artificial intelligence customer service, and remote conferencing, but can also be expanded to multiple fields such as medical image reconstruction and virtual fitting, showing broad application prospects and commercial value.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling, characterized by: The following steps are involved: Initialize the 3D Gaussian model parameters through multi-view images, and divide the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics; Training is done in stages, using a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group; Based on the training results, a 3D digital human model is generated.

2. The semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to claim 1, characterized in that: The 3D Gaussian model parameters are divided into identity group, geometry group and appearance group according to semantics, where The identity group includes position parameters p and scaling matrix S, which are used to maintain the basic identity features of the digital human; The geometry group includes a rotation matrix R and a transparency parameter α, which are used to control the posture and spatial structure; The appearance group includes spherical harmonic coefficients SH, which are used to manage surface lighting and texture information.

3. The semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to claim 1, characterized in that: The stages are divided into a geometric optimization stage and an appearance optimization stage, specifically including: Geometry optimization stage: freeze the 3D Gaussian model parameters of the identity group and appearance group, and update the 3D Gaussian model parameters of the geometry group; Appearance optimization stage: freeze the 3D Gaussian model parameters of the identity group and geometry group, and update the 3D Gaussian model parameters of the appearance group.

4. The semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to claim 3, characterized in that: A local calibration mechanism is introduced in the geometric optimization stage. The triggering conditions of the local calibration mechanism specifically include: Key part offset: The Euclidean distance between the key points of the face or hand and the standard template is greater than the set threshold; Symmetry detection failed: the spatial position difference of the symmetrical structure exceeds the preset threshold; Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds the preset threshold; When one of the trigger conditions of key part offset, symmetry detection failure or posture abnormality detection is met, the local calibration process is activated and the skeleton correction strategy is called to fine-tune the position of the Gaussian points in the corresponding area.

5. The semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to claim 4, characterized in that: The skeleton correction strategy, when the trigger conditions are met, optimizes the mapping function and weights through supervised learning, uses the skeleton correction formula to fine-tune the Gaussian point position in the local area, and resets the geometric group learning rate. The calibration formula is: Among them, P is the original position of the Gaussian point, P corrected is the corrected position, N is the number of joints involved in the transformation, w k is the skin weight of the kth joint, T k (ΔR) is the displacement mapping function corresponding to the kth joint; The displacement mapping function is: T k (ΔR)=W k ·log(ΔR k )+b k Among them, ΔR represents the rotation change between the target posture and the current posture, log(ΔR k ) is the axis-angle representation of the rotation matrix, W k is a learnable linear transformation matrix, b k is the base offset.

6. A semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling, applying the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to any one of claims 1 to 5, characterized in that: include: The grouping module initializes the 3D Gaussian model parameters through multi-view images and divides the 3D Gaussian model parameters into identity group, geometry group and appearance group according to semantics; The training module is used for staged training and uses a differential update mechanism to freeze or update the 3D Gaussian model parameters of each group; The output module is used to generate a 3D digital human model based on the training results.

7. The semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling according to claim 6, characterized in that: The training module also includes: A differential update unit, used to update each set of parameters in stages; A freeze control unit, which manages full freezes of identity groups and staged freezes of other groups; Local calibration unit: used to detect model structure deviations and trigger skeleton correction strategies; Skeleton correction unit: used to perform geometric parameter correction of the formula.

8. The semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling according to claim 6, characterized in that: The grouping module includes identity group, geometry group and appearance group, wherein, The identity group is used to maintain the basic identity characteristics of the digital human by fixing the center position and scaling matrix; Geometry group: used to control the pose and spatial structure by optimizing the rotation matrix and transparency; Appearance group: used to manage surface lighting and texture information by updating spherical harmonic coefficients.

9. The semantic grouping 3D Gaussian sputtering small sample training system for digital human modeling according to claim 7, characterized in that: The local calibration unit is used to detect the following structural anomalies during the geometry optimization phase and serve as a basis for triggering the skeletal correction strategy: Key part offset: The Euclidean distance between the key points of facial features or hands and the standard template is greater than the set threshold; Symmetry detection failed: the spatial position difference of the symmetrical structure exceeds the preset threshold; Posture anomaly detection: the deflection angle of the head or body relative to the frontal view exceeds the preset threshold; When one of the trigger conditions of key part offset, symmetry detection failure or posture abnormality detection is met, the skeleton correction unit is activated to perform the geometric parameter correction operation based on the joint transformation function.

10. A computer storable medium, characterized in that Instructions are stored, and when the instructions are run on a computer, the semantic grouping 3D Gaussian sputtering small sample training method for digital human modeling according to any one of claims 1 to 5 is executed.

Images