A human body gauss point cloud adaptive training system based on a reference attitude geometric constraint

Abstract

The present application provides a kind of human body Gauss point cloud adaptive training system based on reference posture geometric constraint, first, based on the geometric constraint mechanism of reference space inverse projection, by the Gauss point inverse mapping back static reference posture space in training process, and with the distance check of reconstructed surface grid, to inhibit the suspended noise generated by false gradient;Second, based on the smoothing pruning strategy of life cycle cumulative score, use the sum of rendering contribution degree and geometric offset degree of multiple iterations as the discriminant standard, not single frame gradient, improve the time sequence stability of dynamic rendering.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, Gaussian splashing, and human model driving technology, and particularly relates to an adaptive training system for human Gaussian point clouds based on reference posture geometric constraints. Background Technology

[0002] Existing adaptive training methods for human Gaussian point clouds include the following approaches:

[0003] like Figure 1 As shown, 3D Gaussian points are parameterized onto two (front and back) 2D normalized Gaussian maps, and a StyleGAN neural network is used to learn pose-related appearances. Essentially, it transforms the 3D spatial generation problem into a 2D convolutional neural network image generation problem. However, this approach relies on fixed UV mappings or template topology. For large-scale topological changes or geometric deformations caused by movements, it is prone to stretching and undersampling at edges, leading to uneven density.

[0004] like Figure 2 As shown, in pursuit of ultimate rendering speed (166 FPS), this approach pre-defines multiple small MLPs in human body space and performs linear interpolation based on the positions of Gaussian points. However, this approach pre-defines a fixed number of Gaussian points (e.g., 200K) during the initialization phase and drives the process through MLP interpolation during training. Because the number of Gaussian points is statically locked, the system cannot adaptively increase the number of points in detailed areas according to the complexity of human movements (e.g., armpit wrinkles, fluttering clothing).

[0005] like Figure 3 As shown, by combining the SMPL-X full-body parametric model, 3D Gaussian points are directly bound to the triangular surfaces of the mesh model, supporting fine-grained facial and hand actuation. However, this approach causes the Gaussian points, driven by gradients, to easily deviate from the physical surface, resulting in levitation noise and boundary flickering.

[0006] Therefore, the present invention urgently needs to solve the following problems: how to use the baseline pose to establish strong geometric constraints and suppress noise generation in dynamic training; how to achieve adaptive adjustment of point cloud density with action complexity to improve the local detail restoration; and how to automatically synchronize the association weights of the distributed MLP when generating new points to ensure the stability of training convergence. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides an adaptive training system for human Gaussian point clouds based on baseline pose geometric constraints. Through baseline space verification, adaptive densification mechanism, pruning mechanism, and weight recalculation mechanism, it significantly improves the geometric fit of the model, enhances the model's ability to reproduce complex details, maintains an extremely high real-time rendering frame rate, and makes the training process more stable.

[0008] An adaptive training system for human Gaussian point clouds based on reference posture geometric constraints includes a multi-scale reference space construction module, an adaptive evolution execution module, a correlation parameter recalculation module, and a life cycle smoothing pruning module. The multi-scale reference space construction module is used to construct a continuous human body mesh based on multi-view human training videos. and for human body continuous grid Different geometric boundary tolerances are set for Gaussian points at different locations to obtain the set of reference constraint spaces. ; The adaptive evolution execution module is used to select candidate Gaussian points from the Gaussian points in the current round of the set iteration round of the adaptive rendering training of human Gaussian point cloud, map each candidate Gaussian point to the reference coordinate system through LBS inverse transformation, determine whether its position in the reference coordinate system falls into the reference constraint space set Ω, and clone or split the candidate Gaussian points whose judgment result is yes. The correlation parameter recalculation module is used to obtain the interpolation weights of the new Gaussian points obtained by cloning or splitting relative to the spatially distributed MLP architecture. Then, all Gaussian points participate in the adaptive rendering training of subsequent rounds according to their respective interpolation weights and Gaussian properties. The lifecycle smoothing pruning module is used to determine whether the contribution count value of each Gaussian point in the current iteration is greater than a set threshold in the set iteration round of adaptive rendering training of human Gaussian point cloud. Only Gaussian points with a negative result are used to participate in the adaptive rendering training of subsequent rounds. If the contribution of the Gaussian point in the current round is less than the set value, the contribution count value is increased by 1 in the current round. If the contribution of the Gaussian point in the current round is not less than the set value, the contribution count value remains unchanged in the current round.

[0009] Furthermore, determine whether each Gaussian point in the current round falls within the reference constraint space set. The method is as follows: Select the view space position gradient from the Gaussian points in the current round. Greater than the set value The Gaussian points are used as pseudo-dense Gaussian points; The positions of the pseudo-dense Gaussian points in the reference coordinate system were calculated using the inverse LBS transform. Based on the position of each pseudo-dense Gaussian point in the reference coordinate system, determine whether each pseudo-dense Gaussian point falls within the reference constraint space set. .

[0010] Furthermore, the position of any quasi-dense Gaussian point in the reference coordinate system is calculated using the inverse LBS transform as follows:

[0011] in, To determine the position of the pseudo-dense Gaussian point in the reference coordinate system, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh The position in the middle, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh pose parameters in A collection of body parts and regions. For the first Skin weights corresponding to individual body parts. For the first The skeletal transformation matrix corresponding to the individual body part region.

[0012] Furthermore, the method for obtaining the interpolation weights of any new Gaussian point relative to the spatially distributed MLP architecture is as follows: Retrieve the position relative to the current new Gaussian point in a spatially distributed MLP architecture. Recent Distributed MLP nodes, among which... This represents the position of the new Gaussian point in the reference coordinate system after inverse LBS transformation; Based on the above Each distributed MLP node calculates the interpolation weights of the current new Gaussian point relative to each distributed MLP node in the spatial distributed MLP architecture:

[0013] in, The current new Gaussian point is relative to the first point in the spatially distributed MLP architecture. Interpolation weights of distributed MLP nodes Indicates the current position of the new Gaussian point. In the spatially distributed MLP architecture, the first Euclidean distance between distributed MLP nodes; Indicates the new Gaussian point With the retrieved The th distributed MLP node The Euclidean distance between distributed MLP nodes; σ is the distance decay bandwidth parameter that controls the decay rate of the interpolation weights, and the smaller the value of σ, the larger the interpolation weight of the distributed MLP node with a smaller Euclidean distance to the current new Gaussian point.

[0014] Furthermore, the Gaussian properties of any Gaussian point include color, opacity, and position offset, where the position offset is influenced by the continuous human body mesh. Surface constraints.

[0015] Furthermore, the set of reference constraint spaces for:

[0016] in, Let be the coordinates of any position to be tested in the reference constraint space. For x to the continuous grid of the human body The closest distance on the surface, To establish offset tolerances for different regions of the human body, when the location to be tested belongs to the facial region or the finger region... When the location to be examined belongs to the trunk or limbs, .

[0017] Furthermore, the contribution of any Gaussian point is obtained as follows: Obtain the rendering image loss corresponding to the current Gaussian point, and use it as the rendering image loss before masking; Obtain the rendering image loss corresponding to the removal of the current Gaussian point, and use it as the rendering image loss after masking; take the difference ΔL between the rendering image loss before masking and the rendering image loss after masking as the contribution of the current Gaussian point.

[0018] Furthermore, the spatially distributed MLP architecture is used for attitude-driven operation of Gaussian points and outputs the full Gaussian attribute biases of the Gaussian points. ,in, This is the rotation offset. This is the scale offset. This is the transparency offset. This is the color offset.

[0019] Beneficial effects: 1. This invention provides an adaptive training system for human Gaussian point clouds based on reference pose geometric constraints. First, based on the geometric constraint mechanism of inverse projection of the reference space, the Gaussian points in the training process are inversely mapped back to the static reference pose space, and the distance is verified with the reconstructed surface mesh, thereby suppressing the floating noise generated by pseudo gradients. Second, based on the smooth pruning strategy of lifecycle cumulative score, the sum of the rendering contribution and geometric offset of multiple iterations is used as the discrimination criterion, rather than the gradient of a single frame, to improve the temporal stability of dynamic rendering.

[0020] 2. This invention provides an adaptive training system for human Gaussian point clouds based on reference posture geometric constraints, employing a multi-scale, part-specific tolerance strategy. Different geometric boundary ranges Ω are set for different anatomical parts of the human body (such as subtle facial expression areas and large clothing areas on the torso) to achieve a balance between accuracy and robustness.

[0021] 3. This invention provides a human body Gaussian point cloud adaptive training system based on reference posture geometric constraints. According to the joint angular velocity changes of the human body parametric model (SMPL-X), the sensitivity of the densification trigger threshold is increased in advance in the corresponding skin region to achieve density adjustment guided by motion complexity prediction. Attached Figure Description

[0022] Figure 1 A schematic diagram of the prior art provided by the present invention; Figure 2 A schematic diagram of prior art 2 provided by the present invention; Figure 3 A schematic diagram of prior art three provided by the present invention; Figure 4 A schematic diagram of an adaptive training system for human body Gaussian point clouds based on reference posture geometric constraints provided by the present invention; Figure 5 A schematic diagram of the multi-scale geometric constraint benchmark construction module provided by the present invention; Figure 6 This is a schematic diagram of the adaptive evolution execution module provided by the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0024] See Figure 4The system of this invention consists of four core modules: a multi-scale reference space construction module, an adaptive evolution execution module, a correlation parameter recalculation module, and a lifecycle smoothing pruning module. The multi-scale reference space construction module is primarily responsible for MVS reconstruction of T-pose keyframes and Poisson mesh generation, and for defining the geometric boundary tolerances for different parts. The adaptive evolution execution module includes a gradient-based "cloning / splitting" strategy and a validity judgment logic based on benchmark space verification. The correlation parameter recalculation module recalculates the interpolation weights of newly generated points with spatially distributed MLPs. The lifecycle smoothing pruning module is primarily responsible for maintaining the cumulative contribution score of Gaussian points and performing non-instantaneous stability pruning.

[0025] The four core modules are described in detail below.

[0026] I. Multi-scale reference space construction module: like Figure 5 As shown, the multi-scale reference space construction module is used to construct a continuous human body mesh based on multi-view human training videos. and for human body continuous grid Different geometric boundary tolerances are set for Gaussian points at different locations to obtain the set of reference constraint spaces. For example, this invention can extract T-pose frames from multi-view videos, obtain human body surface point clouds through multi-view stereo (MVS), and then generate a continuous human body mesh through Poisson reconstruction. .

[0027] Wherein, the reference constraint space set for:

[0028] in, Let be the coordinates of any position to be tested in the reference constraint space. For x to the continuous grid of the human body The closest distance on the surface, To establish offset tolerances for different regions of the human body, when the location to be tested belongs to the facial region or the finger region... When the location to be examined belongs to the trunk or limbs, .

[0029] It should be noted that, The legal spatial range near the human body surface is defined, not just a single Gaussian point. In the following modules, the reference coordinates of the newly generated Gaussian point after inverse transformation are... Must meet ∈ Only points that meet the criteria are considered geometrically valid and allowed to be added; otherwise, they are considered noise gradients and additions are rejected.

[0030] II. Adaptive Evolution Execution Module like Figure 6 As shown, the adaptive evolution execution module is used to select candidate Gaussian points from the Gaussian points in the current round of the set iteration round of the adaptive rendering training of human Gaussian point cloud, map each candidate Gaussian point to the reference coordinate system through LBS inverse transformation, determine whether its position in the reference coordinate system falls into the reference constraint space set Ω, and clone or split the candidate Gaussian points whose judgment result is yes. For example, in each training session Round iteration (e.g.) If an evolution is triggered, it is determined whether each Gaussian point in the current round falls within the reference constraint space set. The method is as follows: Select the view space position gradient from the Gaussian points in the current round. Greater than the set value The Gaussian points are used as quasi-dense Gaussian points; the positions of the quasi-dense Gaussian points in the reference coordinate system are calculated using the inverse LBS transform; based on the positions of each quasi-dense Gaussian point in the reference coordinate system, it is determined whether each quasi-dense Gaussian point falls into the reference constraint space set. In other words, this invention verifies... The geometric validity is determined by checking if the condition is met. If it is met, a cloning or splitting operation is performed. If it is not met, it is considered an interference noise gradient, and the addition of points is rejected.

[0031] Specifically, the position of any quasi-dense Gaussian point in the reference coordinate system is calculated using the inverse LBS transform as follows:

[0032] in, To determine the position of the pseudo-dense Gaussian point in the reference coordinate system, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh The position in the middle, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh pose parameters in A collection of body parts and regions. For the first Skin weights corresponding to individual body parts. For the first The skeletal transformation matrix corresponding to the individual body part region.

[0033] It should be noted that after the above densification operation is completed, although the newly generated Gaussian points have passed the geometric validity check, they have not yet been associated with the spatially distributed MLP driving network and cannot respond to attitude changes. Therefore, it is necessary to immediately execute the weight synchronization operation of the association parameter recalculation module to assign driving parameters to the new points.

[0034] III. Related Parameter Recalculation Module The correlation parameter recalculation module is used to obtain the interpolation weights of the new Gaussian points obtained by cloning or splitting relative to the spatially distributed MLP architecture. Then, all Gaussian points participate in subsequent rounds of adaptive rendering training according to their respective interpolation weights and Gaussian attributes. The Gaussian attributes of any Gaussian point include color, opacity, and position offset. Color, opacity, and other attributes are directly inherited from the parent point, but the position offset is affected by the human body continuous mesh. The surface constraints ensure that the positional changes remain on the Mesh surface, guaranteeing that the splitting process does not disrupt the surface manifold.

[0035] Furthermore, the method for obtaining the interpolation weights of any new Gaussian point relative to the spatially distributed MLP architecture is as follows: In a spatially distributed MLP architecture, the distance from the current new Gaussian point is retrieved using a KNN. Recent Distributed MLP nodes, among which... This represents the position of the current new Gaussian point in the reference coordinate system after inverse LBS transformation; the preceding KNN retrieval and weight calculation were all performed in the reference space. This is done based on the baseline to ensure that the weight allocation is not affected by the current pose deformation.

[0036] Based on the above Each distributed MLP node calculates the interpolation weights of the current new Gaussian point relative to each distributed MLP node in the spatial distributed MLP architecture:

[0037] in, The current new Gaussian point is relative to the first point in the spatially distributed MLP architecture. Interpolation weights of distributed MLP nodes Indicates the current position of the new Gaussian point. In the spatially distributed MLP architecture, the first Euclidean distance between distributed MLP nodes; Indicates the new Gaussian point With the retrieved The th distributed MLP node The Euclidean distance between distributed MLP nodes; σ is the distance decay bandwidth parameter that controls the decay rate of the interpolation weights, and the smaller the value of σ, the larger the interpolation weight of the distributed MLP node with a smaller Euclidean distance to the current new Gaussian point. The physical meaning of this formula is: the closer the MLP node is to the new point, the greater its driving contribution to the point. Through the exponential decay function and normalization operation, it is ensured that the sum of all weights is 1, thereby achieving smooth spatial interpolation.

[0038] It should be noted that during initialization, based on the Mesh vertex positions, 500 distributed MLP nodes are sampled in a uniform distribution. Each MLP generates weights for its surrounding Gaussian points. Each Gaussian point is controlled by its K nearest neighbor MLPs based on distance. This spatially distributed MLP architecture is used to drive the pose of Gaussian points and output the full Gaussian attribute biases of the Gaussian points. ,in, This is the rotation offset. This is the scale offset. This is the transparency offset. This is the color offset.

[0039] The correlation parameter recalculation module is a key module connecting the adaptive evolution execution module (densening) and the lifecycle smoothing pruning module (pruning). The new Gaussian points generated by the adaptive evolution execution module through cloning or splitting only possess geometric position information and lack the ability to be driven by pose parameters. The correlation parameter recalculation module employs a spatially distributed MLP architecture for pose driving, where each Gaussian point needs to hold a set of MLP interpolation weights to participate in rendering. Therefore, after a new Gaussian point is generated in the reference space, the system must immediately establish a driving association between it and surrounding MLP nodes.

[0040] At this point, after geometric verification by the adaptive evolution execution module and weight synchronization by the associated parameter recalculation module, the newly generated Gaussian points possess both valid spatial locations and complete driving parameters, and can formally participate in subsequent rendering training. However, as training progresses, some Gaussian points may gradually lose their rendering contribution due to viewpoint occlusion or pose changes, requiring cleanup through the pruning mechanism of the lifecycle smoothing pruning module.

[0041] IV. Lifecycle Smooth Pruning Module The lifecycle smoothing pruning module is used to determine whether the contribution count value of each Gaussian point in the current iteration is greater than a set threshold in the set iteration round of adaptive rendering training of human Gaussian point cloud. Only Gaussian points with a negative result are used to participate in the adaptive rendering training of subsequent rounds. If the contribution of the Gaussian point in the current round is less than the set value, the contribution count value is increased by 1 in the current round. If the contribution of the Gaussian point in the current round is not less than the set value, the contribution count value remains unchanged in the current round.

[0042] It should be noted that after the adaptive evolution execution module's densification and the correlation parameter recalculation module's weight synchronization, the number of Gaussian points in the system will continue to increase. To prevent redundant points from accumulating and causing excessive computational overhead, this step introduces a lifecycle counter to record the continuous contribution status of each Gaussian point. For example, in the early stages of training, initial pruning is performed every 2000 steps and ends at 30,000 steps; within every 2000 steps, the contribution of all Gaussian points is calculated every 50 steps. Specifically, the method for obtaining the contribution of any Gaussian point is as follows: Obtain the rendering image loss corresponding to the current Gaussian point, and use it as the rendering image loss before masking; Obtain the rendering image loss corresponding to the removal of the current Gaussian point, and use it as the rendering image loss after masking; take the difference ΔL between the rendering image loss before masking and the rendering image loss after masking as the contribution of the current Gaussian point.

[0043] For example, from the current perspective, calculate the image loss before and after masking a certain Gaussian point. The difference ,like ,but If not satisfied, If it remains unchanged, it will not be pruned; every 2000 steps, all Gaussian point scores are reset, and the next round of pruning begins; that is, it only when... If the action is invalid for 20 consecutive rounds, pruning is performed. This avoids accidental deletion due to single-frame occlusion and ensures the continuity of the action sequence.

[0044] In summary, compared to unconstrained 3DGS solutions, this invention reduces the "detachment" phenomenon of Gaussian points under violent movements by approximately 92% through benchmark space verification, solves the boundary flickering problem, and significantly improves geometric fit. The adaptive densification mechanism increases the point cloud density in areas such as clothing wrinkles and facial expressions by approximately 2 times, enhancing the ability to reproduce complex details. The pruning mechanism can accurately remove redundant points, reducing the total number of Gaussian points by 25%-30% while maintaining the same rendering quality, thereby maintaining an extremely high real-time rendering frame rate (160+ FPS) and making computational resources more efficiently utilized. The weight recalculation mechanism solves the "attribute mutation" problem after adding new points, improves the model convergence speed by more than 15%, reduces local collapse phenomena, and makes the training process more stable.

[0045] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. An adaptive training system for human body Gaussian point clouds based on reference posture geometric constraints, characterized in that, It includes a multi-scale baseline space construction module, an adaptive evolution execution module, a correlation parameter recalculation module, and a lifecycle smoothing pruning module; The multi-scale reference space construction module is used to construct a continuous human body mesh based on multi-view human training videos. and for human body continuous grid Different geometric boundary tolerances are set for Gaussian points at different locations to obtain the set of reference constraint spaces. ; The adaptive evolution execution module is used to select candidate Gaussian points from the Gaussian points in the current round of the set iteration round of the adaptive rendering training of human Gaussian point cloud, map each candidate Gaussian point to the reference coordinate system through LBS inverse transformation, determine whether its position in the reference coordinate system falls into the reference constraint space set Ω, and clone or split the candidate Gaussian points whose judgment result is yes. The correlation parameter recalculation module is used to obtain the interpolation weights of the new Gaussian points obtained by cloning or splitting relative to the spatially distributed MLP architecture. Then, all Gaussian points participate in the adaptive rendering training of subsequent rounds according to their respective interpolation weights and Gaussian properties. The lifecycle smoothing pruning module is used to determine whether the contribution count value of each Gaussian point in the current iteration is greater than a set threshold in the set iteration round of adaptive rendering training of human Gaussian point cloud. Only Gaussian points with a negative result are used to participate in the adaptive rendering training of subsequent rounds. If the contribution of the Gaussian point in the current round is less than the set value, the contribution count value is increased by 1 in the current round. If the contribution of the Gaussian point in the current round is not less than the set value, the contribution count value remains unchanged in the current round.

2. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, Determine whether each Gaussian point in the current round falls within the reference constraint space set. The method is as follows: Select the view space position gradient from the Gaussian points in the current round. Greater than the set value The Gaussian points are used as pseudo-dense Gaussian points; The positions of the pseudo-dense Gaussian points in the reference coordinate system were calculated using the inverse LBS transform. Based on the position of each pseudo-dense Gaussian point in the reference coordinate system, determine whether each pseudo-dense Gaussian point falls within the reference constraint space set. .

3. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 2, characterized in that, The specific steps for calculating the position of any quasi-dense Gaussian point in the reference coordinate system using the LBS inverse transform are as follows: in, To determine the position of the pseudo-dense Gaussian point in the reference coordinate system, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh The position in the middle, To achieve pseudo-dense Gaussian points in the current round of the human body continuous mesh pose parameters in A collection of body parts and regions. For the first Skin weights corresponding to individual body parts. For the first The skeletal transformation matrix corresponding to the individual body part region.

4. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, The method for obtaining the interpolation weights of any new Gaussian point relative to the spatially distributed MLP architecture is as follows: Retrieve the position relative to the current new Gaussian point in a spatially distributed MLP architecture. Recent Distributed MLP nodes, among which... This represents the position of the new Gaussian point in the reference coordinate system after inverse LBS transformation; Based on the above Each distributed MLP node calculates the interpolation weights of the current new Gaussian point relative to each distributed MLP node in the spatial distributed MLP architecture: in, The current new Gaussian point is relative to the first point in the spatially distributed MLP architecture. Interpolation weights of distributed MLP nodes Indicates the current position of the new Gaussian point. In the spatially distributed MLP architecture, the first Euclidean distance between distributed MLP nodes; Indicates the new Gaussian point With the retrieved The th distributed MLP node The Euclidean distance between distributed MLP nodes; σ is the distance decay bandwidth parameter that controls the decay rate of the interpolation weights, and the smaller the value of σ, the larger the interpolation weight of the distributed MLP node with a smaller Euclidean distance to the current new Gaussian point.

5. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, The Gaussian properties of any Gaussian point include color, opacity, and position offset, where the position offset is influenced by the continuous human body mesh. Surface constraints.

6. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, The set of reference constraint spaces for: in, Let be the coordinates of any position to be tested in the reference constraint space. For x to the continuous grid of the human body The closest distance on the surface, To establish offset tolerances for different regions of the human body, when the location to be tested belongs to the facial region or the finger region... When the location to be examined belongs to the trunk or limbs, .

7. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, The contribution of any Gaussian point is obtained as follows: Obtain the rendering image loss corresponding to the current Gaussian point, and use it as the rendering image loss before masking; Obtain the rendering image loss corresponding to the removal of the current Gaussian point, and use it as the rendering image loss after masking; take the difference ΔL between the rendering image loss before masking and the rendering image loss after masking as the contribution of the current Gaussian point.

8. The human body Gaussian point cloud adaptive training system based on reference posture geometric constraints as described in claim 1, characterized in that, The spatially distributed MLP architecture is used to perform attitude control on Gaussian points and output the full Gaussian attribute biases of the Gaussian points. ,in, This is the rotation offset. This is the scale offset. This is the transparency offset. This is the color offset.

Images