Three-dimensional editing method and system based on anchor point visual angle guidance and three-dimensional perception attention pooling

Abstract

The invention discloses a three-dimensional editing method and system based on anchor point visual angle guidance and three-dimensional perception attention pooling. According to the method, a framework combining a diffusion model and three-dimensional Gaussian rendering is designed and is used for improving the consistency and efficiency of multi-view editing. According to the method, firstly, an anchor point view angle guiding strategy is put forward, editing is carried out through cross-view angle attention on a small number of key view angles, and an editing result is spread to other view angles by utilizing depth information, so that cross-view angle calculation overhead is reduced, and consistency is kept. Then, a three-dimensional perception attention pooling mechanism is constructed, text-image attention results of different visual angles are back-projected to unified three-dimensional Gaussian representation, and global consistent semantic attention distribution is realized through cross-visual-angle aggregation so as to enhance three-dimensional structure perception of a diffusion model. And finally, the three-dimensional model is directly updated by generating a multi-view editing result at one time, so that a redundant iterative optimization process in a traditional method is avoided.

Description

Technical Field

[0001] This invention relates to the field of 3D editing technology, and in particular to a 3D editing method and system based on anchor point perspective guidance and 3D perception attention pooling. Background Technology

[0002] In recent years, with the increasing demand for applications such as virtual reality, digital twins, film and television production, and intelligent design, 3D generation and editing technologies have received widespread attention. Emerging methods based on 3D Gaussian Splatting and diffusion models have made it possible to generate 3D content driven by text or images. However, existing methods still have significant shortcomings in practical applications: on the one hand, traditional 3D editing relies heavily on global optimization, lacking targeted control over local details and key semantic regions, resulting in editing results that often fail to accurately reflect user intent; on the other hand, consistency in multi-view editing is difficult to guarantee, and editing effects from different perspectives may exhibit geometric distortion, texture drift, or semantic deviation, severely affecting the spatial perception and realism of 3D models. Especially in editing tasks involving complex scenes or intricate objects, existing technologies struggle to balance editing flexibility with spatial consistency. Therefore, there is an urgent need for a 3D editing method that can introduce anchor point perspectives as guidance, combining spatial perception and multi-view consistency to provide higher-quality technical support for 3D content generation and interaction. Summary of the Invention

[0003] This invention addresses the shortcomings of existing 3D generation and editing methods in terms of local semantic control and multi-view consistency by proposing a 3D editing method and system based on anchor point view guidance and 3D perception attention pooling to improve semantic accuracy and cross-view stability.

[0004] According to the first aspect of this specification, a 3D editing method based on anchor point view guidance and 3D perceptual attention pooling is provided. This method, through the collaboration of 3D Gaussian Splatting (3DGS) and a diffusion model, performs single-batch editing of multi-view rendered images and directly updates the 3D representation. The method includes: S1: Model the target 3D scene using 3D Gaussian representation, generate the 3D Gaussian representation to be edited using multi-view rendering, and obtain multi-view images and corresponding depth maps by rendering with preset camera poses; S2: Edit the input text conditions and the multi-view image input diffusion model; S3: In each sampling of the diffusion sampling, select several anchor point views from the view set as a reference for cross-view information fusion and fine editing; S4: For the anchor point perspective, cross-perspective self-attention is introduced in the self-attention calculation of the diffusion model denoising process, so that the query vector of the anchor point perspective pays attention to the features of this perspective and other perspectives at the same time, so as to achieve consistent semantic editing across perspectives. S5: For non-anchor point views, geometric constraints are applied in the diffusion model denoising process based on the depth of the nearest anchor point view in 3D space and the camera extrinsic parameters: the anchor point view pixels are back-projected to 3D points, and then the 3D points are projected to the target view (the non-anchor point view) to limit the candidate region. The matching position is obtained by optimizing according to the feature cosine similarity, and the target view features and the anchor point view matching features are linearly fused according to the weighted coefficient, thereby completing the geometric-based semantic propagation. S6: The diffusion model denoising process performs 3D perceptual attention pooling: calculate text-image cross-attention for each viewpoint, map the text-image cross-attention of each viewpoint according to transparency and transmittance and converge it onto the 3D Gaussian point cloud representation, perform weighted pooling on the point cloud in 3D space and then project it back to each viewpoint to obtain a consistent attention distribution across viewpoints to constrain the diffusion model's editing of the image. S7: The diffusion model is decoded to obtain the editing results from each perspective, and the 3D Gaussian point cloud representation is jointly optimized and updated in a single forward pass using the editing results.

[0005] Furthermore, in step S3, the selection of anchor point viewpoints is dynamically rotated: within multiple time steps of diffusion sampling, the viewpoint set is first traversed in a non-repeating manner. When all views have been traversed and the time step has not yet ended, the process is repeated; at least one traversal is completed so that each viewpoint in the viewpoint set is used as an anchor point viewpoint at least once.

[0006] Furthermore, in step S4, the cross-view self-attention of the anchor point perspective adopts a weighted fusion of self-attention and cross-view attention, and the fusion formula is as follows: ; in, The feature representation representing the current target perspective (or query) Feature representations of other perspectives (or contexts) being referenced; , , These are linear projection matrices of query, key, and value, respectively. Scaling factor Number of reference viewpoints; weight parameters ;when Increasing the size improves the structural fidelity of this viewpoint. Reduce the cross-view appearance alignment capability; the obtained cross-view self-attention value Replace the original self-attention value in the diffusion model.

[0007] Furthermore, in step S5, the search range is limited by geometric constraints, which are calculated as follows: (a) First, the pixels in the anchor point view. Through depth With camera parameters Backprojection to 3D points: ; in, Represents the horizontal and vertical coordinates of pixels. These are the camera intrinsic parameters, rotation matrix, and spatial position, respectively. (b) Then, the three-dimensional points are adjusted according to the camera parameters of the target viewpoint. Projecting onto the target viewpoint yields the initial matching position: ; in, Represents the perspective projection operator; (c) This yields the target viewpoint candidate region: ; in, For the search radius, Pixels in the target's viewpoint.

[0008] Furthermore, in step S5, the candidate region is limited by a threshold, which is the search radius. , The spatial position of the anchor point view and the target view, camera intrinsic parameters, rotation matrix, and depth variance are adaptively set to expand the search radius when there is significant occlusion and parallax.

[0009] Furthermore, in step S5, feature localization uses cosine similarity metric to select the most similar pixel position from the candidate region corresponding to the anchor point viewpoint. .

[0010] Furthermore, in step S5, the target viewpoint features Features matching anchor point view The fusion is performed using the following formula: ; in, The weighting coefficients are adaptively calculated based on the spatial position of the anchor point view and the target view, camera intrinsic parameters, rotation matrix, and reprojection error.

[0011] Furthermore, when S4 and S5 are combined, the method performs cross-view self-attention only on the anchor point view, while using geometrically guided feature propagation on the non-anchor point view, thereby significantly reducing the computational and memory overhead of cross-view attention.

[0012] Further, the 3D perceptual attention pooling in step S6 includes the following process: in the text-image cross-attention computation of the diffusion model denoising process, 3D perceptual attention pooling is performed at each layer or a specified layer; specifically, for each Gaussian kernel... First, it aggregates the text-image cross-attention representations in the set of visible views. , i Representing the i One perspective, u Represents pixel position, l Representing textual conditions, thereby obtaining an attention representation in three-dimensional space. Then, the attention representation in the three-dimensional space is projected back to each viewpoint to obtain a cross-viewpoint consistent attention representation. To replace the original cross-attention representation Aggregate weights include Gaussian opacity. With line of sight transmittance The calculation formula is: ; ; in, Represents text conditions. For the set of visible viewpoints, This indicates a projection operation.

[0013] Furthermore, the diffusion model is a diffusion-based image editing model that follows natural language instructions, such as a diffusion model based on the InstructPix2Pix model. The embedded representation obtained by the text conditional text encoder is used to drive self / cross-attention and the 3D perceptual attention pooling.

[0014] Furthermore, in step S7, the 3D update is performed using the paradigm of "single forward-multiple view simultaneous editing-joint backhaul", which eliminates the need to edit each view individually and rebuild in multiple rounds, thereby shortening the inference time and reducing iteration redundancy.

[0015] According to a second aspect of this specification, a 3D editing system based on anchor point perspective guidance and 3D perception attention pooling is provided, the system comprising: The rendering module is used to generate multi-view images and corresponding depth maps from 3DGS representation; Anchor point selection module, used to dynamically select anchor point viewpoints during diffusion sampling; The anchor point editing module is used to perform cross-view self-attention semantic editing on the anchor point perspective; The geometry propagation module is used to define candidate regions based on depth and camera extrinsics from a non-anchor point perspective and to perform feature matching and fusion. The 3D attention pooling module is used to aggregate text-image data from multiple perspectives across attention in the 3D Gaussian domain and backfeed it to each perspective. The reconstruction and update module is used to optimize and update the 3D Gaussian point cloud representation based on the edited multi-view images.

[0016] Compared with the prior art, the beneficial effects of this invention are as follows: 1. This invention proposes an anchor-point-guided 3D editing framework that can provide spatial awareness information through anchor-point perspectives, thereby effectively alleviating inconsistency issues in multi-view editing.

[0017] 2. This invention designs a spatial awareness attention mechanism that transmits the geometric and semantic information of the anchor point perspective to other perspectives, thereby achieving cross-perspective feature alignment and information fusion.

[0018] 3. This invention proposes a geometric consistency constraint, which enhances the geometric stability of cross-view editing by guiding the anchor point perspective, thus avoiding the local drift and deformation distortion problems that occur in traditional methods.

[0019] 4. Experiments on various 3D editing tasks show that the present invention can significantly improve multi-view consistency while maintaining semantic accuracy, verifying the effectiveness and superiority of the method. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall framework provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of anchor point-guided editing provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of three-dimensional perception attention aggregation provided in an embodiment of the present invention; Figure 4 These are comparison diagrams of traditional 3D editing methods provided in the embodiments of the present invention; Figure 5 This is a 3D editing effect diagram of the surrounding scene provided in an embodiment of the present invention; Figure 6 This is a 3D editing effect diagram of the forward scene provided in an embodiment of the present invention. Detailed Implementation

[0022] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0023] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0024] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0025] This invention provides a 3D editing method based on anchor point perspective guidance and 3D perception attention pooling. This method not only significantly improves editing efficiency but also outperforms existing technologies in terms of semantic consistency and spatial structure preservation. The method includes the following steps: S1: Model the target 3D scene using 3D Gaussian representation, generate the 3D Gaussian representation to be edited using multi-view rendering, and obtain multi-view images and corresponding depth maps by rendering with preset camera poses; S2: Edit the input text conditions and the multi-view image input diffusion model; S3: In each sampling of the diffusion sampling, select several anchor point views from the view set as a reference for cross-view information fusion and fine editing; In one implementation, the diffusion model in step S2 is based on the InstructPix2Pix model; the selection of anchor viewpoints in step S3 is dynamically rotated: within multiple time steps of diffusion sampling, the viewpoint set is first traversed in a non-repeating manner. When all views have been traversed and the time step has not yet ended, the process is repeated; at least one traversal is completed so that each viewpoint in the viewpoint set serves as an anchor viewpoint at least once. For example, starting from the first time step, each time step selects the same number of views from the viewpoint set as anchor viewpoints without repetition. When all views in the viewpoint set have been selected and the time step has not yet ended, the same number of views are selected again from the original viewpoint set without repetition as anchor viewpoints until the time step ends. Simultaneously, the diffusion model is a diffusion-based image editing model that follows natural language instructions, and the embedding obtained by the text condition through the text encoder is used to drive self / cross-attention and the 3D perceptual attention pooling.

[0026] S4: For the anchor point perspective, cross-perspective self-attention is introduced in the self-attention calculation of the diffusion model denoising process, so that the query vector of the anchor point perspective pays attention to the features of this perspective and other perspectives at the same time, so as to achieve consistent semantic editing across perspectives. In one implementation, the cross-view self-attention of the anchor point perspective in step S4 adopts a weighted fusion of self-attention and cross-view attention, and the fusion formula is as follows: ; in, The feature representation representing the current target perspective (or query) Feature representations of other perspectives (or contexts) being referenced; , , These are linear projection matrices of query, key, and value, respectively. Scaling factor Number of reference viewpoints; weight parameters ;when Increasing the size improves the structural fidelity of this viewpoint. This reduces the cross-view appearance alignment capability. The resulting cross-view self-attention value is then used to enhance this capability. Replace the original self-attention value in the diffusion model.

[0027] S5: For non-anchor point views, geometric constraints are applied in the diffusion model denoising process based on the depth of the nearest anchor point view in 3D space and the camera extrinsic parameters: the anchor point view pixels are back-projected to 3D points, and then the 3D points are projected to the target view (the non-anchor point view) to limit the candidate region. The matching position is obtained by optimizing according to the feature cosine similarity, and the target view features and the anchor point view matching features are linearly fused according to the weighted coefficient, thereby completing the geometric-based semantic propagation. In one implementation, the search range in step S5 is defined by geometric constraints, which are calculated as follows: (a) First, the pixels in the anchor point view. Through depth With camera parameters Backprojection to 3D points: ; in, Represents the horizontal and vertical coordinates of pixels. These are the camera intrinsic parameters, rotation matrix, and spatial position, respectively. (b) Then, the three-dimensional points are adjusted according to the camera parameters of the target viewpoint. Projecting onto the target viewpoint yields the initial matching position: ; in, Represents the perspective projection operator; (c) This yields the target viewpoint candidate region: ; in, For the search radius, Pixels in the target's viewpoint.

[0028] In one implementation, the candidate region in step S5 is defined by a threshold, which is the search radius. , The spatial positions of the anchor point view and the target view, camera intrinsics, rotation matrix, and depth variance are adaptively set to expand the search radius when occlusion and parallax are significant. Feature localization uses cosine similarity metric to select the most similar pixel position corresponding to the anchor point view from the candidate region. Target view features in step S5 Features matching anchor point view The fusion is performed using the following formula: ; in, The weighting coefficients are adaptively calculated based on the spatial position of the anchor point view and the target view, camera intrinsic parameters, rotation matrix, and reprojection error.

[0029] When combining steps S4 and S5, this method performs cross-view self-attention only on the anchor point view, while using geometrically guided feature propagation on non-anchor point views, thereby significantly reducing the computational and memory overhead of cross-view attention.

[0030] S6: The diffusion model denoising process performs 3D perceptual attention pooling: calculate text-image cross-attention for each viewpoint, map the text-image cross-attention of each viewpoint according to transparency and transmittance and converge it onto the 3D Gaussian point cloud representation, perform weighted pooling on the point cloud in 3D space and then project it back to each viewpoint to obtain a consistent attention distribution across viewpoints to constrain the diffusion model's editing of the image. In one implementation, the 3D perceptual attention pooling in step S6 includes the following process: performing 3D perceptual attention pooling at each layer or a specified layer during the text-image cross-attention computation in the diffusion model denoising process; specifically, for each Gaussian kernel First, it aggregates the text-image cross-attention representations in the set of visible views. ( i Representing the i One perspective, u Represents pixel position, l (representing textual conditions), thereby obtaining an attention representation in three-dimensional space. Then, the attention representation in the three-dimensional space is projected back to each viewpoint to obtain a cross-viewpoint consistent attention representation. To replace the original cross-attention representation Aggregate weights include Gaussian opacity. With line of sight transmittance The calculation formula is: ; ; in, Represents text conditions. For the set of visible viewpoints, This indicates a projection operation.

[0031] S7: The diffusion model is decoded to obtain the editing results from each perspective, and the 3D Gaussian point cloud representation is jointly optimized and updated in a single forward pass using the editing results.

[0032] In one implementation, the 3D update in step S7 is performed in the paradigm of "single forward-multiple view simultaneous editing-joint backhaul", which does not require editing each view individually and rebuilding in multiple rounds, thereby shortening the inference time and reducing iteration redundancy.

[0033] This method proposes an anchor-point-guided 3D editing approach that combines 3D perceptual attention pooling with geometric consistency constraints to improve the consistency and semantic accuracy of multi-view editing. The anchor-point-guided and 3D perceptual attention pooling 3D editing method used in this embodiment is described below with reference to the accompanying figures.

[0034] Figure 1 This demonstrates the overall framework of the method of this invention. First, input text prompts and an initial 3D Gaussian representation are used. Then, the anchor point selection module selects the reference viewpoint that best reflects the semantics. Next, semantic consistency editing is performed on this viewpoint, and the results are gradually propagated to other viewpoints through rendering and optimization. The entire process includes text guidance, semantic hotspot detection, multi-view consistency control, and 3D updates, embodying the idea of ​​"from a single anchor point to global consistency."

[0035] Figure 2 The diagram illustrates anchor-guided editing. On the left, semantic relevance is calculated from text prompts in multi-view renderings, and the anchor view with the best semantics is selected. On the right, it shows how to use the features of the anchor to constrain and propagate to other views, ensuring that the global 3D representation remains consistent with the semantic goal as it is progressively updated.

[0036] Figure 3This diagram illustrates 3D attention aggregation. Features from different perspectives are first aligned through semantic attention, and then aggregated using a 3D spatial awareness strategy to integrate key cross-perspective information into a global 3D representation. Compared to simple image feature fusion, this part explicitly models 3D geometric relationships, thus ensuring consistency at both semantic and geometric levels.

[0037] Figure 4 The diagram shows a comparison of traditional 3D editing methods, highlighting the differences between previous methods that relied on per-view editing and multiple reconstructions to gradually ensure consistency and the proposed method that generates semantically consistent multi-view editing in one go based on a diffusion model with 3D perception.

[0038] Figure 5 The comparison of final editing results from different methods in a surrounding scene is shown. Traditional methods often result in blurriness, incomplete textures, or semantic deviations from certain perspectives, while this method generates clear, structurally sound, and semantically accurate results from all angles, demonstrating its advantage in global consistency.

[0039] Figure 6 The presentation demonstrates the effectiveness of this method in forward scene editing tasks. Compared to traditional methods, this method, after editing from the anchor point perspective, better maintains structural consistency and semantic accuracy across multiple perspectives, avoiding issues such as misalignment and distortion, thus illustrating the effectiveness of this method.

[0040] Corresponding to the aforementioned embodiments of the 3D editing method based on anchor point perspective guidance and 3D perceptual attention pooling, the present invention also provides embodiments of a 3D editing system for implementing the above method, the system comprising: The rendering module is used to generate multi-view images and corresponding depth maps from a 3D Gaussian point cloud representation; Anchor point selection module, used to dynamically select anchor point viewpoints during diffusion sampling; The anchor point editing module is used to perform cross-view self-attention semantic editing on the anchor point perspective; The geometry propagation module is used to define candidate regions based on depth and camera extrinsics from a non-anchor point perspective and to perform feature matching and fusion. The 3D attention pooling module is used to aggregate text-image data from multiple perspectives across attention in the 3D Gaussian domain and backfeed it to each perspective. The reconstruction and update module is used to optimize and update the 3D Gaussian point cloud representation based on the edited multi-view images.

[0041] Corresponding to the aforementioned embodiments of the 3D editing method based on anchor point perspective guidance and 3D perception attention pooling, the present invention also provides embodiments of a 3D editing device based on anchor point perspective guidance and 3D perception attention pooling.

[0042] The 3D editing device provided in this embodiment of the invention includes a memory and one or more processors. The memory stores executable code. When the processor executes the executable code, it is used to implement the 3D editing method based on anchor point view guidance and 3D perception attention pooling in the above embodiment.

[0043] The embodiments of the 3D editing device based on anchor point perspective guidance and 3D perception attention pooling of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar equipment. The device embodiments can be implemented through software, hardware, or a combination of both.

[0044] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the 3D editing method based on anchor point perspective guidance and 3D perception attention pooling as described in the above embodiments.

[0045] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0046] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of one or more embodiments of this specification.

Claims

1. A 3D editing method based on anchor point perspective guidance and 3D perception attention pooling, characterized in that, include: S1: Model the target 3D scene using 3D Gaussian representation, generate the 3D Gaussian representation to be edited using multi-view rendering, and obtain multi-view images and corresponding depth maps by rendering with preset camera poses; S2: Edit the input text conditions and the multi-view image input diffusion model; S3: In each sampling of the diffusion sampling, select several anchor point views from the view set as a reference for cross-view information fusion and fine editing; S4: For the anchor point perspective, cross-perspective self-attention is introduced in the self-attention calculation of the diffusion model denoising process, so that the query vector of the anchor point perspective pays attention to the features of this perspective and other perspectives at the same time, so as to achieve consistent semantic editing across perspectives. S5: For non-anchor point views, geometric constraints are applied in the diffusion model denoising process based on the depth of the nearest anchor point view in 3D space and the camera extrinsic parameters: the anchor point view pixels are back-projected to 3D points, and then the 3D points are projected to the target view to limit the candidate region. The matching position is obtained by optimizing according to the feature cosine similarity, and the target view features and the anchor point view matching features are linearly fused according to the weighted coefficient, thereby completing the geometry-based semantic propagation. S6: The diffusion model denoising process performs 3D perceptual attention pooling: calculate text-image cross-attention for each viewpoint, map the text-image cross-attention of each viewpoint according to transparency and transmittance and converge it onto the 3D Gaussian point cloud representation, perform weighted pooling on the point cloud in 3D space and then project it back to each viewpoint to obtain a consistent attention distribution across viewpoints to constrain the diffusion model's editing of the image. S7: The diffusion model is decoded to obtain the editing results from each perspective, and the 3D Gaussian point cloud representation is jointly optimized and updated in a single forward pass using the editing results.

2. The method according to claim 1, characterized in that, In step S3, the selection of the anchor point viewpoint is dynamically rotated: within multiple time steps of diffusion sampling, the viewpoint set is first traversed in a non-repeating manner. When all views have been traversed and the time step has not yet ended, the process is repeated; at least one traversal is completed so that each viewpoint in the viewpoint set is used as an anchor point viewpoint at least once.

3. The method according to claim 1, characterized in that, In step S4, the cross-view self-attention of the anchor point perspective adopts a weighted fusion of self-attention and cross-view attention, and the fusion formula is as follows: ； in, This represents the feature representation of the current target perspective. This represents the feature representation of other perspectives being referenced; , , These are the linear projection matrices of query, key, and value, respectively. Scaling factor Number of reference viewpoints; weight parameters ;when Increasing the size improves the structural fidelity of this viewpoint. Reduce the cross-view appearance alignment capability; the obtained cross-view self-attention value Replace the original self-attention value in the diffusion model.

4. The method according to claim 1, characterized in that, In step S5, the search range is limited by geometric constraints, which are calculated as follows: (a) Pixels in the anchor point view Through depth With camera parameters Backprojection to 3D points: ； in, Represents the horizontal and vertical coordinates of pixels. These are the camera intrinsic parameters, rotation matrix, and spatial position, respectively. (b) The three-dimensional points are adjusted according to the camera parameters of the target viewpoint. Projecting onto the target viewpoint yields the initial matching position: ； in, Represents the perspective projection operator; (c) Obtain the target view candidate region: ； in, For the search radius, Pixels in the target's viewpoint.

5. The method according to claim 4, characterized in that, In step S5, the candidate region is limited by a threshold, which is the search radius. , The spatial position of the anchor point view and the target view, camera intrinsic parameters, rotation matrix, and depth variance are adaptively set to expand the search radius when there is significant occlusion and parallax.

6. The method according to claim 1, characterized in that, In step S5, target viewpoint features Features matching anchor point view The fusion is performed using the following formula: ； in, The weighting coefficients are adaptively calculated based on the spatial position of the anchor point view and the target view, camera intrinsic parameters, rotation matrix, and reprojection error.

7. The method according to claim 1, characterized in that, Step S6, the 3D perceptual attention pooling, includes the following process: In the text-image cross-attention computation of the diffusion model denoising process, 3D perceptual attention pooling is performed at each layer or a specified layer; specifically, for each Gaussian kernel... First, it aggregates the text-image cross-attention representations in the set of visible views. , i Representing the i One perspective, u Represents pixel position, l Representing textual conditions, thereby obtaining an attention representation in three-dimensional space. Then, the attention representation in the three-dimensional space is projected back to each viewpoint to obtain a cross-viewpoint consistent attention representation. To replace the original cross-attention representation Aggregate weights include Gaussian opacity. With line of sight transmittance The calculation formula is: ； ； in, For the set of visible viewpoints, This indicates a projection operation.

8. The method according to claim 1, characterized in that, The diffusion model is a diffusion-based image editing model that follows natural language instructions. The embedded representation of the text conditions obtained by the text encoder is used to drive self / cross-attention and the 3D perceptual attention pooling.

9. The method according to claim 1, characterized in that, In step S7, the 3D update is performed using the paradigm of "single forward pass - simultaneous editing of multiple views - joint back pass". This eliminates the need to edit each view individually and rebuild in multiple rounds, thereby shortening the inference time and reducing iteration redundancy.

10. A three-dimensional editing system for implementing the method according to any one of claims 1-9, characterized in that, include: The rendering module is used to generate multi-view images and corresponding depth maps from a 3D Gaussian point cloud representation; Anchor point selection module, used to dynamically select anchor point viewpoints during diffusion sampling; The anchor point editing module is used to perform cross-view self-attention semantic editing on the anchor point perspective; The geometry propagation module is used to define candidate regions based on depth and camera extrinsics from a non-anchor point perspective and to perform feature matching and fusion. The 3D attention pooling module is used to aggregate text-image data from multiple perspectives across attention in the 3D Gaussian domain and backfeed it to each perspective. The reconstruction and update module is used to optimize and update the 3D Gaussian point cloud representation based on the edited multi-view images.

Images