A Gaussian Space-Guided Interactive Two-Handed 3D Reconstruction Method
By introducing a Gaussian space guidance mechanism, the spatial relationship between the two hands is explicitly depicted, which solves the problem of insufficient stability of the monocular RGB hand 3D reconstruction method in interactive two-hand scenarios and achieves higher accuracy and consistency in 3D reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI'AN POLYTECHNIC UNIVERSITY
- Filing Date
- 2026-05-19
- Publication Date
- 2026-06-30
AI Technical Summary
Existing monocular RGB hand 3D reconstruction methods lack stability in interactive two-hand scenarios, making it difficult to accurately depict the spatial relationship between the hands. In particular, semantic confusion and geometric distortion are prone to occur under occlusion and contact conditions.
A spatial geometry guidance mechanism based on pre-trained Gaussian fields is introduced to explicitly characterize the spatial relationship between the hands during feature modeling. Combined with multi-scale feature fusion and cross-modal feature fusion, the stability and accuracy of the network for 3D reconstruction of the hands are improved.
It significantly improves the structural integrity and spatial consistency of interactive two-handed 3D reconstruction, reduces semantic confusion and geometric distortion, and is suitable for stable reconstruction in complex scenarios.
Smart Images
Figure CN122312922A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of computer vision, 3D geometric modeling and hand pose estimation, and specifically relates to an interactive two-hand 3D reconstruction method based on Gaussian space guidance. Background Technology
[0002] Hand 3D reconstruction technology has significant application value in fields such as natural human-computer interaction, virtual reality / augmented reality, robot teaching, and gesture understanding. With the popularization of mobile terminals and lightweight AR / VR devices, the demand for low-cost, easy-to-deploy interactive two-hand 3D reconstruction solutions that do not require depth sensors and rely solely on a monocular RGB camera is becoming increasingly urgent. However, in interactive two-hand scenarios, due to the high similarity in appearance between the left and right hands and frequent occlusion and contact, traditional methods that rely solely on the appearance features of two-dimensional images are difficult to accurately depict the spatial relationship between the hands. This can easily lead to blurred structures in occluded areas, geometric distortion, or semantic confusion between the left and right hands, thus seriously affecting the integrity and spatial consistency of the reconstruction results.
[0003] Existing monocular RGB hand 3D reconstruction methods mainly include keypoint detection-based methods and parametric hand model-based methods. The former describes hand pose information by predicting 2D or 3D keypoints, but it is difficult to recover the complete 3D geometric surface; the latter regresses hand pose and shape parameters and generates a 3D mesh by combining a predefined hand mesh model to achieve end-to-end reconstruction. To further improve reconstruction accuracy, existing methods have introduced strategies such as multi-scale feature fusion and attention mechanisms in feature extraction and network structure, but they still rely heavily on 2D image appearance features and lack explicit modeling and stable constraints on the 3D spatial geometry, resulting in insufficient reconstruction stability in scenarios with hand interaction and severe occlusion. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, the purpose of this invention is to provide an interactive two-handed 3D reconstruction method based on Gaussian space guidance. By introducing a spatial geometric guidance mechanism based on pre-trained Gaussian fields during the feature modeling process, the network's ability to understand the spatial distribution and 3D relationship of the hands is enhanced under monocular RGB input conditions. This enables the reconstruction process, which originally relied solely on 2D visual features, to acquire spatial perception capabilities. Thus, without relying on multi-view imaging or depth sensors, the stability and accuracy of interactive two-handed 3D reconstruction in complex scenes are effectively improved.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A Gaussian space-guided interactive two-handed 3D reconstruction method, characterized by the following steps:
[0007] Step 1: Data preprocessing, 2D feature extraction, and early separation of left-hand and right-hand features:
[0008] Obtain a monocular RGB image containing the interacting hands: The RGB image is preprocessed, including hand region detection and alignment, image cropping to a uniform size, and pixel value normalization. The preprocessed image is then input into a ResNet-50-based backbone feature encoding network, which extracts multi-scale two-dimensional visual features. During feature extraction, early separation of left and right hand features is achieved through 1×1 convolution on the highest-level feature map. This process can be represented as follows:
[0009] ;
[0010] Where H, W, and C represent the height, width, and number of channels of the feature map, respectively. and Initial feature representations for the left and right hands are obtained respectively, resulting in independent initial features for the left and right hands; by establishing independent representation spaces for the left and right hands at the feature layer, the subsequent feature enhancement and relation modeling processes can be facilitated.
[0011] Step 2, Hands Feature Enhancement Module:
[0012] The initial features of the left and right hands are input into the bimanual feature enhancement module to perform structural enhancement and relationship modeling on the left and right hand features, explicitly establishing the interaction relationship constraints between the left and right hands, and obtaining a bimanual feature representation with enhanced structural consistency. Specifically, this includes the following two sub-steps:
[0013] (1) Cross-hand relationship modeling mechanism;
[0014] This process employs a cross-hand relationship modeling mechanism to exchange information between left and right hand features, explicitly characterizing the relationship between the two hands in interactive scenarios. This process utilizes left and right hand features... and Using the input as input, a bidirectional cross-hand feature mapping relationship is established, enabling the features of one hand to perceive the spatial distribution and structural state of the other hand. This process can be represented as follows:
[0015] ;
[0016] in, This indicates a cross-hand relationship modeling operation. and These represent the right-hand and left-hand features obtained after introducing information from the other hand, respectively, which are used to establish geometric and semantic constraints between the left and right hands at the feature layer.
[0017] (2) Enhanced hierarchical features;
[0018] Based on the completed cross-hand relationship modeling, hierarchical enhancement processing is performed on the features of both hands. By progressively enhancing and fusing features at different scales and semantic levels, a more structurally consistent and discriminative representation of the features of both hands is obtained. The process can be represented as follows:
[0019] ;
[0020] in, This represents a hierarchical feature enhancement operation. and These are the enhanced features of the left and right hands, respectively.
[0021] Step 3: Gaussian space guidance;
[0022] Introducing a 3D Gaussian geometric field obtained through multi-view data pre-training during the independent phase. The three-dimensional Gaussian geometric field is used to characterize the continuous geometric distribution characteristics of the hand in three-dimensional space. It is frozen during the training and inference of the main network and does not participate in end-to-end parameter updates. The three-dimensional Gaussian geometric field is projected onto the image plane through differentiable rendering to obtain two-dimensional Gaussian features that are spatially aligned with the two-dimensional visual features. These features are used as spatial geometric guidance information. The two-dimensional Gaussian features are represented as follows:
[0023] ;in, (⋅) represents a differentiable rendering or Gaussian sputtering projection process. and These represent spatial resolutions consistent with two-dimensional visual features. The number of channels for the Gaussian feature;
[0024] Step 4: Cross-modal feature fusion;
[0025] The enhanced hand features obtained in the second step Compared with the two-dimensional Gaussian features generated in the third step The input cross-modal feature fusion module generates fused bifurcated features containing spatial geometric constraint information through feature alignment and weighted fusion.
[0026] ;
[0027] in (⋅) represents a cross-modal fusion operation, which achieves the fusion of two-dimensional visual features and spatial geometric features through gating weighting and feature alignment, thereby obtaining a biphasic feature representation containing spatial information. ;
[0028] Step 5: 3D hand model regression and generation;
[0029] The fused hand features are input into a 3D hand model regression module to regress the pose and shape parameters of the left and right hands. Based on these parameters, corresponding 3D mesh models of the left and right hands are generated, thereby achieving interactive 3D reconstruction of both hands under monocular RGB image conditions. The fused features are then used to... Input regressor to predict left and right hand MANO parameters:
[0030] ;
[0031] in (⋅) represents a parametric regression network. These are the left and right hand posture parameters, respectively. These are the shape parameters for the left and right hands, respectively; the parameters are input into the MANO model to generate 3D meshes for the left and right hands. The final interactive two-handed 3D reconstruction result was obtained.
[0032] The early separation process of left and right hand features is carried out in the two-dimensional visual feature encoding stage. By using convolutional mapping, independent representation spaces for the left and right hands are established in the feature layer, so that the left and right hand features remain distinguishable to each other in subsequent processing, thereby reducing the risk of semantic confusion under conditions of similar appearance of the two hands and occlusion.
[0033] The aforementioned two-hand feature enhancement module includes two sub-processes: cross-hand relationship modeling and hierarchical feature enhancement. Cross-hand relationship modeling is used to characterize the spatial relationship between the left and right hands, while hierarchical feature enhancement is used to progressively enhance hand features at different scales and semantic levels to improve the expressive power of interactive areas and key structural parts.
[0034] The three-dimensional Gaussian geometric field is obtained through independent pre-training with multi-view data outside the main network. Its parameters are frozen during the training and inference process of the main network, and only the generated two-dimensional Gaussian spatial features are used as spatial geometric priors to participate in feature modeling, thereby ensuring that the overall reconstruction process still strictly meets the monocular RGB input conditions.
[0035] The two-dimensional Gaussian features are normalized and feature-mapped, then projected onto a high-dimensional geometric feature space to obtain a spatially guided feature representation for subsequent modulation and fusion.
[0036]
[0037] in, This represents a feature projection or embedding mapping operation. The spatial saliency distribution characterized by Gaussian geometric priors is preserved. Through the above processing, the two-dimensional Gaussian features generated based on the pre-trained Gaussian field are used as spatial guidance information to participate in the subsequent feature fusion and pose estimation processes, so as to provide stable spatial geometric constraints.
[0038] The cross-modal feature fusion module fuses two-dimensional visual features with two-dimensional Gaussian spatial features through gating weighting or feature alignment, enabling spatial geometric priors to adaptively participate in feature modeling at different spatial locations, thereby providing stable spatial constraints in occluded or hand-clasped areas.
[0039] The parametric hand model is a model that generates a three-dimensional mesh based on hand posture parameters and shape parameters. Through this model, the local structural morphology of the left and right hands and their overall configuration relationship in three-dimensional space can be recovered simultaneously, thereby obtaining a structurally complete and spatially consistent interactive two-hand three-dimensional reconstruction result.
[0040] The beneficial effects of this invention are:
[0041] This invention effectively compensates for the lack of spatial information under monocular RGB conditions by using Gaussian space guidance, significantly reducing semantic confusion and geometric distortion between the left and right hands, and improving the performance of interactive two-handed 3D reconstruction in terms of structural integrity, spatial consistency, and robustness. It has good versatility and practical application value. Practical use has verified that this method has good generalization ability and is suitable for various complex two-handed interaction scenarios. Attached Figure Description
[0042] Figure 1 A flowchart illustrating an example of the present invention;
[0043] Figure 2 This is a schematic diagram of the method model of the present invention;
[0044] Figure 3 This is a schematic diagram of the structure of the hand feature enhancement module;
[0045] Figure 4 This is a schematic diagram of the Gaussian space guidance module structure proposed in this invention;
[0046] Figure 5 This is a schematic diagram of the cross-modal feature fusion structure designed in this invention;
[0047] Figure 6 The hand pose estimation results of this invention on the InterHand2.6M and FreiHand datasets are visualized. Detailed Implementation
[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0049] like Figure 1 As shown, Figure 1 This is a flowchart of a monocular RGB interactive two-handed 3D reconstruction method based on Gaussian space representation, according to an embodiment of the present invention. Based on a prepared dataset, stable reconstruction of the interactive two-handed 3D structure is achieved.
[0050] S1. Data Preparation and Preprocessing (Main Network Data Stream and Gaussian Field Data Stream)
[0051] In one embodiment, the system first acquires a monocular RGB image containing the interacting hands. The image is then preprocessed. These preprocessing operations include, but are not limited to: hand region detection and alignment, image cropping to a preset size, and pixel value normalization to eliminate differences in scale, position, and brightness between different input images.
[0052] During the data preparation phase, the preprocessed RGB image serves as the input to the main network for subsequent 2D visual feature extraction and 3D reconstruction. Simultaneously, the system introduces a set of pre-trained offline 3D Gaussian geometric field data. This Gaussian geometric field originates from the independent training process of multi-view data and remains frozen during the main network training and inference phases, serving as a spatial geometric prior. By decoupling the main network data stream and the Gaussian field data stream in subsequent stages, the overall reconstruction process strictly adheres to the monocular RGB input conditions.
[0053] S2. Construct a monocular RGB interactive biphasic 3D reconstruction network based on Gaussian space representation.
[0054] In one embodiment, a neural network structure for interactive two-handed 3D reconstruction is constructed, such as... Figure 2 As shown. The network adopts an end-to-end processing approach and includes at least a two-dimensional visual feature encoding module, a left-hand and right-hand feature early separation module, a two-hand feature enhancement module, a Gaussian space guidance module, a cross-modal feature fusion module, and a three-dimensional hand model regression module. These modules are connected sequentially and work together.
[0055] The preprocessed RGB image is first input into a two-dimensional visual feature encoding module, where multi-scale two-dimensional visual feature representations F are extracted through convolutional layers and residual structures. During the feature extraction stage, a 1×1 convolution operator is introduced onto the high-level feature maps to achieve early separation of left-hand and right-hand features, obtaining the initial left-hand features. With right-hand initial features This allows for the establishment of independent representation spaces for the left and right hands at the feature layer, reducing the risk of left- and right-hand feature confusion in subsequent interaction scenarios.
[0056] Subsequently, the initial features of the left and right hands are input into the bi-hand feature enhancement module, the structure of which is shown in the diagram below. Figure 3As shown, the two-hand feature enhancement module is used to jointly model and enhance the features of the left and right hands. By explicitly modeling the relationship between the two hands, the features of one hand can perceive the spatial distribution and structural state of the other hand, thereby improving the overall consistency of the features of both hands in interactive scenarios. At the same time, this module performs hierarchical enhancement of hand features to strengthen the expressive ability of key structures such as fingertips, palm edges, and contact areas of both hands, obtaining a more discriminative representation of the features of both hands.
[0057] After enhancing the features of both hands, a Gaussian space guidance module is introduced, the structure of which is shown in the diagram below. Figure 4 As shown, the Gaussian spatial guidance module generates two-dimensional Gaussian spatial features aligned with the two-dimensional visual features in terms of spatial resolution, based on a pre-trained three-dimensional Gaussian geometric field. These features characterize the continuous geometric distribution of the hand in three-dimensional space. This spatial feature serves as an external geometric prior, guiding the spatial modeling of visual features within the network, enabling the network to maintain stable spatial perception capabilities even in occluded or hand-clasped areas.
[0058] Furthermore, the visual features enhanced by the hand features and the Gaussian spatial features are input into the cross-modal feature fusion module, the structure of which is shown in the diagram below. Figure 5 As shown, the cross-modal feature fusion module fuses two-dimensional visual features with spatial geometric features through gating weighting and feature alignment, generating a two-handed feature representation containing spatial information. Through this fusion method, Gaussian spatial priors can adaptively participate in feature modeling at different spatial locations, thereby introducing stable spatial geometric constraints without compromising visual semantic expression.
[0059] Finally, the fused hand features are fed into the 3D hand model regression module for subsequent prediction of pose and shape parameters.
[0060] S3. Data input network training (main network training and Gaussian field pre-training)
[0061] In one embodiment, the three-dimensional Gaussian geometric field is pre-trained in an independent phase using multi-view training data before the main network is trained. This Gaussian field is used to learn the continuous geometric distribution characteristics of the hand in three-dimensional space. After pre-training, the parameters are frozen and do not participate in the backpropagation update of the main network.
[0062] During the main network training phase, the preprocessed monocular RGB image is input into the constructed interactive two-handed 3D reconstruction network, and forward propagation is performed to obtain the intermediate feature representations, fused features, and corresponding pose and shape prediction results for the left and right hands. Since Gaussian space features are introduced as static geometric priors, they only participate in feature modulation and fusion during training and do not affect the main network's modeling settings for monocular inputs.
[0063] S4. Training Optimization Based on Joint Loss Function
[0064] In one embodiment, to effectively supervise the interactive two-handed 3D reconstruction process, a joint loss function is used to constrain the network training process. The joint loss function consists of multiple losses, which are weighted and summed to form an overall optimization objective, in the form of:
[0065]
[0066] in, This represents the hand model reconstruction loss, used to constrain the predicted hand joint positions and 3D mesh geometry; This represents the loss of the relative spatial relationship between the two hands, used to constrain the three-dimensional relative translational relationship between the left and right hands; The pixel-wise supervision loss represents intermediate features and is used to improve the stability of feature representation. These are the weighting coefficients for the corresponding loss terms.
[0067] By backpropagating the joint loss function and iteratively updating the main network parameters, the network gradually learns the feature representation ability that takes into account both two-dimensional appearance information and three-dimensional spatial structure. The pre-trained Gaussian field is kept frozen during the training process and only participates in feature modeling as a spatial geometric prior.
[0068] S5, 3D hand model regression and hand mesh generation
[0069] The fused hand features obtained in the preceding steps are input into the 3D hand model regression module to estimate the pose and shape parameters of the left and right hands, and generate corresponding 3D mesh models of the left and right hands based on these parameters. By parametrically constraining the geometric constraints of the hand models, the local structural morphology of both hands and their overall spatial configuration can be recovered simultaneously, thereby obtaining a structurally complete and spatially consistent interactive 3D reconstruction result.
[0070] The above embodiments illustrate a specific implementation of the present invention in a monocular RGB interactive two-handed 3D reconstruction task from the perspective of system architecture design and processing flow organization. These embodiments are only used to illustrate the technical concept and feasible implementation path of the present invention and are not intended to limit the scope of protection of the present invention. In actual engineering applications, the composition of network modules, feature processing strategies, the introduction method of Gaussian space priors, or data interaction processes can be appropriately adjusted or optimized according to specific deployment environments, computing resource conditions, real-time requirements, or differences in application scenarios. As long as the relevant technical solutions are still based on the core technical idea of the present invention—introducing 3D spatial geometric information through Gaussian space representation and co-modeling with 2D visual features—and do not substantially change their overall working mechanism and technical effects, they should all be considered reasonable engineering implementations of the technical solutions of the present invention and fall within the scope of protection of the present invention.
Claims
1. A method for interactive dual-hand 3D reconstruction based on Gaussian space guidance, characterized in that, Includes the following steps: Step 1: Data preprocessing, 2D feature extraction, and early separation of left-hand and right-hand features: Acquiring monocular RGB images containing interacting hands: ; The RGB image is preprocessed, including hand region detection and alignment, image cropping to a uniform size, and pixel value normalization. The preprocessed image is then input into a ResNet-50-based backbone feature encoding network, which extracts multi-scale two-dimensional visual features. During the feature extraction stage, 1×1 convolutions are used on the highest-level feature map to achieve early separation of left and right hand features. This process can be represented as follows: wherein H, W, C represent the height, width and channel number of the feature map respectively, and are the initial feature representations of the left hand and the right hand respectively, obtaining the left hand initial feature and the right hand initial feature which are independent of each other. By establishing independent representation spaces for the left and right hands at the feature layer, the subsequent feature enhancement and relation modeling processes can be facilitated. Step 2, Hands Feature Enhancement Module: The initial features of the left and right hands are input into the bimanual feature enhancement module to perform structural enhancement and relationship modeling on the left and right hand features, explicitly establishing the interaction relationship constraints between the left and right hands, and obtaining a bimanual feature representation with enhanced structural consistency. Specifically, this includes the following two sub-steps: (1) Cross-hand relationship modeling mechanism; This process employs a cross-hand relationship modeling mechanism to exchange information between left and right hand features, explicitly characterizing the relationship between the two hands in interactive scenarios. This process utilizes left and right hand features... and Using the input as input, a bidirectional cross-hand feature mapping relationship is established, enabling the features of one hand to perceive the spatial distribution and structural state of the other hand. This process can be represented as follows: ； in, This indicates a cross-hand relationship modeling operation. and These represent the right-hand and left-hand features obtained after incorporating information from the other hand, respectively, which are used to establish geometric and semantic constraints between the left and right hands at the feature layer; (2) Enhanced hierarchical features; Based on the completed cross-hand relationship modeling, hierarchical enhancement processing is performed on the features of both hands. By progressively enhancing and fusing features at different scales and semantic levels, a more structurally consistent and discriminative representation of the features of both hands is obtained. The process can be represented as follows: ； in, This represents a hierarchical feature enhancement operation. and These are the enhanced features of the left and right hands, respectively. Step 3: Gaussian space guidance; Introducing a 3D Gaussian geometric field obtained through multi-view data pre-training during the independent phase. The three-dimensional Gaussian geometric field is used to characterize the continuous geometric distribution characteristics of the hand in three-dimensional space. It is frozen during the training and inference of the main network and does not participate in end-to-end parameter updates. The three-dimensional Gaussian geometric field is projected onto the image plane through differentiable rendering to obtain two-dimensional Gaussian features that are spatially aligned with the two-dimensional visual features. These features are used as spatial geometric guidance information. The two-dimensional Gaussian features are represented as follows: ;in, (⋅) represents a differentiable rendering or Gaussian sputtering projection process. and These represent spatial resolutions consistent with two-dimensional visual features. The number of channels for the Gaussian feature; Step 4: Cross-modal feature fusion; The enhanced hand features obtained in the second step Compared with the two-dimensional Gaussian features generated in the third step The input cross-modal feature fusion module generates fused bifurcated features containing spatial geometric constraint information through feature alignment and weighted fusion. ； in (⋅) represents a cross-modal fusion operation, which achieves the fusion of two-dimensional visual features and spatial geometric features through gating weighting and feature alignment, thereby obtaining a biphasic feature representation containing spatial information. ; Step 5: 3D hand model regression and generation; The fused hand features are input into a 3D hand model regression module to regress the pose and shape parameters of the left and right hands. Based on these parameters, corresponding 3D mesh models of the left and right hands are generated, thereby achieving interactive 3D reconstruction of both hands under monocular RGB image conditions. The fused features are then used to... Input regressor to predict left and right hand MANO parameters: ； in (⋅) represents a parametric regression network. These are the left and right hand posture parameters, respectively. These are the shape parameters for the left and right hands, respectively; the parameters are input into the MANO model to generate 3D meshes for the left and right hands. The final interactive two-handed 3D reconstruction result was obtained.
2. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The early separation process of left and right hand features is carried out in the two-dimensional visual feature encoding stage. By using convolutional mapping, independent representation spaces for the left and right hands are established in the feature layer, so that the left and right hand features remain distinguishable to each other in subsequent processing, thereby reducing the risk of semantic confusion under conditions of similar appearance of the two hands and occlusion.
3. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The aforementioned two-hand feature enhancement module includes two sub-processes: cross-hand relationship modeling and hierarchical feature enhancement. Cross-hand relationship modeling is used to characterize the spatial relationship between the left and right hands, while hierarchical feature enhancement is used to progressively enhance hand features at different scales and semantic levels to improve the expressive power of interactive areas and key structural parts.
4. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The three-dimensional Gaussian geometric field is obtained through independent pre-training with multi-view data outside the main network. Its parameters are frozen during the training and inference process of the main network, and only the generated two-dimensional Gaussian spatial features are used as spatial geometric priors to participate in feature modeling, thereby ensuring that the overall reconstruction process still strictly meets the monocular RGB input conditions.
5. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The two-dimensional Gaussian features are normalized and feature-mapped, then projected onto a high-dimensional geometric feature space to obtain a spatially guided feature representation for subsequent modulation and fusion. in, This represents a feature projection or embedding mapping operation. The spatial saliency distribution characterized by Gaussian geometric priors is preserved. Through processing, the two-dimensional Gaussian features generated based on the pre-trained Gaussian field are used as spatial guidance information to participate in the subsequent feature fusion and pose estimation process, so as to provide stable spatial geometric constraints.
6. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The cross-modal feature fusion module fuses two-dimensional visual features with two-dimensional Gaussian spatial features through gating weighting or feature alignment, enabling spatial geometric priors to adaptively participate in feature modeling at different spatial locations, thereby providing stable spatial constraints in occluded or hand-clasped areas.
7. The interactive two-handed 3D reconstruction method based on Gaussian space guidance according to claim 1, characterized in that, The parametric hand model is a model that generates a three-dimensional mesh based on hand posture parameters and shape parameters. Through this model, the local structural morphology of the left and right hands and their overall configuration relationship in three-dimensional space can be recovered simultaneously, thereby obtaining a structurally complete and spatially consistent interactive two-hand three-dimensional reconstruction result.