Three-dimensional mesh generation method based on surface element autoregressive compression and electronic device

Abstract

The application provides a three-dimensional mesh generation method based on patch autoregressive compression and an electronic device. The three-dimensional mesh generation method based on patch autoregressive compression comprises the following steps: converting an input three-dimensional shape representation into an ordered triangular patch sequence, wherein each triangular patch is defined by the coordinates of its three vertices; using a shape encoder to encode the input three-dimensional shape representation into a latent representation; using an autoregressive decoder to generate each triangular patch in the ordered triangular patch sequence one by one in sequence, with the latent representation as a condition, thereby generating a three-dimensional mesh. By representing a three-dimensional mesh as an ordered patch sequence and performing autoregressive generation, the modeling granularity is improved from the bottom vertex coordinates to the high-level triangular patch. The model can directly learn and generate the topological structure of the mesh, significantly improving the ability to capture the geometric structure and topological relationship of the mesh, thereby achieving the purpose of improving the generation quality and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a three-dimensional mesh generation method and electronic device based on surface element autoregressive compression. Background Technology

[0002] 3D content generation (AIGC 3D) is a core task in computer graphics and vision, widely used in games, virtual reality (VR), industrial design, and other fields. Among various 3D representations, triangular meshes have become the industry standard due to their high rendering efficiency and ability to explicitly represent complex topological structures. With the development of deep learning, direct mesh generation has become a hot topic.

[0003] Existing autoregressive mesh generation methods (such as MeshGPT) typically tile the mesh as a sequence of vertex coordinates. For example, a triangle face contains 3 vertices, each with 3 coordinates, requiring a total of 9 tokens. This results in extremely long sequences, and because the Transformer's self-attention mechanism has a quadratic complexity, the computational and memory consumption increases dramatically as the number of mesh faces increases, making it difficult to generate high-resolution meshes containing thousands of faces. Existing compression methods (such as block indexing or complex traversal) often sacrifice topological connectivity or introduce a huge vocabulary, and block indexing-based methods can lead to vocabulary explosion, increasing training difficulty. Existing methods operate at a low semantic level of "vertex coordinates." However, the basic topological unit of a mesh is a triangle face, not isolated coordinates. This low-level tokenization makes it difficult for the model to capture high-level geometric structures and topological relationships. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a three-dimensional mesh generation method and electronic device based on surface element autoregressive compression, which at least partially solves the problem of difficulty in capturing high-level geometric structures and topological relationships in the prior art.

[0005] In a first aspect, embodiments of this disclosure provide a three-dimensional mesh generation method based on surface element autoregressive compression, including:

[0006] The input 3D shape representation is converted into an ordered sequence of triangle faces, where each triangle face is defined by the coordinates of its three vertices;

[0007] The input 3D shape representation is encoded into a latent representation using a shape encoder;

[0008] Using an autoregressive decoder, each triangle face in the ordered sequence of triangle faces is generated sequentially, conditioned on the latent representation, to generate a three-dimensional mesh.

[0009] Optionally, generating each triangle face in the ordered sequence of triangle faces sequentially, conditioned on the latent representation, includes:

[0010] Each triangle face is treated as an independent and indivisible semantic unit, such that the length of the generated sequence of the entire mesh is equal to the number of triangle faces it contains.

[0011] Optionally, the input 3D shape representation is converted into an ordered sequence of triangle faces, including:

[0012] Determine which point each triangular facet represents in a 3D shape;

[0013] Based on the spatial coordinates of the representative points of all triangle faces, all triangle faces are sorted according to a predetermined lexicographical order to form an ordered sequence of triangle faces.

[0014] Optionally, determining that each triangular facet in the three-dimensional shape represents a point includes:

[0015] The vertex with the smallest coordinate value among the three vertices of the triangular face is selected as the representative point of the face.

[0016] Optionally, the step of using an autoregressive decoder to sequentially generate each triangle face in the ordered triangle face sequence, conditioned on the latent representation, includes:

[0017] For the triangle face to be predicted, at least one triangle face that has been generated before is mapped to a corresponding face token by the face embedding layer.

[0018] The face token sequence is input into the Transformer decoder, and the latent features of the current face are output through a causal self-attention mechanism and a cross-attention mechanism conditioned on the latent representation.

[0019] The decoding head decodes the latent features of the current face into the coordinates of the three vertices.

[0020] Optionally, decoding the latent features of the current face into the coordinate values ​​of three specific vertices using the decoding head includes:

[0021] The CausalMLP is used as the decoding head. When predicting multiple coordinates within a triangle face, the CausalMLP employs an in-plane autoregressive mechanism, meaning that the prediction of subsequent coordinates depends on the previously predicted prior coordinates within the same face.

[0022] Optionally, the input three-dimensional shape representation includes point clouds, voxels, or multi-view images.

[0023] Optionally, the shape encoder is a VecSet-based encoder used to convert the input point cloud into a set of latent vectors to obtain a latent representation.

[0024] Optionally, a shape encoder is used to encode the input 3D shape representation into a latent representation, including:

[0025] Sample k query points from the input point cloud.

[0026] The query point is used as the Query, and the original complete point cloud as the Key and Value. Global geometric information is aggregated to generate initial features. ;

[0027] Initial features After processing by the Transformer encoder, the latent representation C is output.

[0028] Secondly, embodiments of this disclosure also provide an electronic device, the electronic device comprising:

[0029] At least one processor; and,

[0030] A memory communicatively connected to the at least one processor; wherein,

[0031] The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the three-dimensional mesh generation method based on surface element autoregressive compression as described in any of the first aspects.

[0032] The present invention provides a 3D mesh generation method based on autoregressive compression of facets. By representing the 3D mesh as an ordered sequence of facets and generating it through autoregression, the modeling granularity is increased from the bottom-level vertex coordinates to the high-level triangular facets. This enables the model to directly learn and generate the mesh topology, significantly improving the ability to capture the mesh geometry and topological relationships, thereby improving both generation quality and efficiency. Attached Figure Description

[0033] The above and other objects, features and advantages of this disclosure will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0034] Figure 1 A schematic diagram of an asymmetric encoder-decoder architecture provided in an embodiment of this disclosure;

[0035] Figure 2 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0036] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0037] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0038] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0039] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The illustrations only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0040] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0041] This embodiment discloses a three-dimensional mesh generation method based on surface element autoregressive compression, including:

[0042] The input 3D shape representation is converted into an ordered sequence of triangle faces, where each triangle face is defined by the coordinates of its three vertices;

[0043] The input 3D shape representation is encoded into a latent representation using a shape encoder;

[0044] Using an autoregressive decoder, each triangle face in the ordered sequence of triangle faces is generated sequentially, conditioned on the latent representation, to generate a three-dimensional mesh.

[0045] The input 3D shape representation is converted into an ordered sequence of triangle faces, including:

[0046] Determine which point each triangular facet represents in a 3D shape;

[0047] Based on the spatial coordinates of the representative points of all triangle faces, all triangle faces are sorted according to a predetermined lexicographical order to form an ordered sequence of triangle faces.

[0048] This embodiment abandons complex traversal algorithms and adopts a simple and deterministic space sorting strategy.

[0049] Specific method: For each face in the grid, take the vertex with the smallest coordinates among its three vertices as the representative point of that face.

[0050] Sort the points lexicographically based on their coordinates, that is, compare Z first, then Y, and finally X.

[0051] This sorting method provides a standardized order, eliminates system complexity, and experiments have shown that it delivers a lower Hausdorff distance error than DFS / BFS.

[0052] Based on the latent representation, each triangle face in the ordered sequence of triangle faces is generated sequentially, including:

[0053] Each triangle face is treated as an independent and indivisible semantic unit, such that the length of the generated sequence of the entire mesh is equal to the number of triangle faces it contains.

[0054] Using an autoregressive decoder, each triangle face in the ordered sequence of triangle faces is generated sequentially, conditioned on the latent representation, including:

[0055] For the triangle face to be predicted, at least one triangle face that has been generated before is mapped to a corresponding face token by the face embedding layer.

[0056] The face token sequence is input into the Transformer decoder, and the latent features of the current face are output through a causal self-attention mechanism and a cross-attention mechanism conditioned on the latent representation.

[0057] The decoding head decodes the latent features of the current face into the coordinates of the three vertices.

[0058] This refers to the "One-Face-One-Token" strategy.

[0059] For each face in the sequence, it is defined by the coordinates of three vertices, which flatten out to be a 9-dimensional vector. Unlike existing technologies that tokenize these 9 values ​​separately, this embodiment treats the entire 9-dimensional vector as a single unit. A dedicated embedding / pooling layer projects this 9-dimensional vector into a d-dimensional token, which is then directly input into the Transformer.

[0060] The sequence length was reduced from 9F to F by using a one-word-per-side strategy, achieving a compression ratio of 0.11.

[0061] Determine which triangle face in a 3D shape represents a point, including:

[0062] The vertex with the smallest coordinate value among the three vertices of the triangular face is selected as the representative point of the face.

[0063] The decoding head decodes the latent features of the current face into the coordinate values ​​of three specific vertices, including:

[0064] The CausalMLP is used as the decoding head. When predicting multiple coordinates within a triangle face, the CausalMLP employs an in-plane autoregressive mechanism, meaning that the prediction of subsequent coordinates depends on the previously predicted prior coordinates within the same face.

[0065] The input is a 3D shape representation, including point clouds, voxels, or multi-view images.

[0066] The shape encoder is a VecSet-based encoder used to convert an input point cloud into a set of latent vectors to obtain a latent representation.

[0067] like Figure 1 As shown, this embodiment adopts an asymmetric encoder-decoder architecture.

[0068] VecSet Shape Encoder

[0069] The encoder's role is to compress the input point cloud P into a compact latent representation C.

[0070] Architecture selection: 3DShape2VecSet architecture is adopted.

[0071] Processing flow:

[0072] 1. Furthest Point Sampling (FPS): Sampling k query points from the input point cloud.

[0073] 2. Cross Attention: These query points serve as the Query, and the original complete point cloud serves as the Key and Value. Global geometric information is aggregated to generate initial features. The formula is as follows:

[0074] ,

[0075] 3. Transformer Encoder: Initial Features After processing by the L_E layer Transformer encoder, the final latent representation C (Latent VecSet) is output.

[0076] Autoregressive Face Decoder:

[0077] The decoder uses the latent representation C to generate a mesh sequence face by face. This embodiment abandons the externally pre-trained VQ-VAE and adopts an end-to-end training method to learn the internal surface representation.

[0078] The decoding process involves three key steps:

[0079] Step 1: Face Embedding

[0080] Use the face generated in the previous step as input.

[0081] A lightweight MLP (called Face Pooling Layer) is used to map 9-dimensional continuous coordinates to a single d-dimensional token.

[0082] It realizes the transformation from physical space to potential space.

[0083] Step 2: Transformer Processing

[0084] The token sequence enters the Transformer decoder stack layer.

[0085] Each layer contains two attention operations:

[0086] 1. Causal Self-Attention: Captures the local structure and connectivity of the grid. It can only see tokens from the previous step.

[0087] ,

[0088] 2. Cross-Attention: Injects global shape context. The self-attention output is used as the query, and the encoder's VecSet C is used as the key and value.

[0089] ,

[0090] The output is the latent vector of the i-th face.

[0091] Step 3: Face Decoding Head (CausalMLP)

[0092] A key component for reconstructing latent vectors into geometric coordinates. This embodiment designs a CausalMLP head that introduces an in-plane second-order autoregressive mechanism.

[0093] Mechanism: In order to predict 9 coordinates in a plane, the prediction of the j-th coordinate depends not only on the latent vector output by the Transformer, but also on the coordinates that have been predicted in the plane before.

[0094] Compared to parallel prediction of 9 coordinates or decoding using attention layers, CausalMLP can more effectively maintain the geometric consistency of vertices within the face.

[0095] Output: The probability distribution (Logits) of the discretized tokens for each coordinate.

[0096] Training Objective

[0097] The model is trained end-to-end to minimize reconstruction loss.

[0098] The coordinates are quantized (e.g., quantized to integers [0,127] or [0,1023]).

[0099] The loss function is the average of the standard cross-entropy loss across all coordinates on all faces.

[0100] This embodiment also demonstrates the potential of this architecture as a foundational model for 3D generation.

[0101] Latent Diffusion Model: Train a Diffusion Transformer (DiT) conditioned on 2D image features (from DINOv3) to generate the VecSet latent representation C disclosed in this embodiment.

[0102] The input image is processed by a diffusion model, and the predicted VecSet is processed by a FACE decoder with frozen weights to obtain a high-fidelity mesh.

[0103] This demonstrates that the latent space learned by FACE has high semantic robustness.

[0104] The one-face-one-token strategy in this embodiment treats each triangle face (containing the coordinate data of its three vertices) that makes up the mesh as an independent, indivisible token unit, rather than treating each coordinate or each vertex as a token. In autoregressive models, this strategy reduces the mesh sequence length from the order of vertex coordinates to the order of faces (a compression ratio of approximately 1:9).

[0105] The Face Embedding layer is used to directly project a 9-dimensional face coordinate vector into a potential token that can be processed by the Transformer.

[0106] When the CausalMLP decoding head decodes the latent face token back to 9 coordinates, it adopts an in-face autoregression mechanism, that is, the prediction conditions of subsequent coordinates depend on the prediction results of the prior coordinates in the same face.

[0107] A VecSet-based shape encoder and the aforementioned face-based autoregressive decoder are used to directly generate triangular meshes with explicit topological structures from point clouds or latent space.

[0108] The present invention is compared with the prior art as follows:

[0109] 1. Extreme sequence compression efficiency (Compression Ratio):

[0110] MeshGPT has a compression ratio of 0.67, MeshAnythingV2 has a compression ratio of 0.28, and EdgeRunner has a compression ratio of 0.46.

[0111] This invention: FACE achieves a compression ratio of 0.11, which is more than twice that of existing technologies. This means that for the same complex mesh, the sequence length generated by FACE is only about 1 / 6 of that of MeshGPT.

[0112] 2. Significantly reduce computing costs:

[0113] Existing technologies are limited by squared attention complexity, resulting in slow inference and high memory consumption.

[0114] This invention reduces sequence length by 9 times, theoretically reduces the computational cost of the self-attention layer by 81 times, and reduces GPU memory usage by approximately 9 times. This makes it possible to generate higher resolution (more facets) meshes with limited resources.

[0115] 3. SOTA Rebuild Quality:

[0116] Achieved the best reconstruction quality on the Objaverse and Toys4K datasets.

[0117] The method disclosed in this embodiment has the following advantages:

[0118] 1. Improve efficiency: By using the one-face-one-token strategy, each triangle face is treated as a single token, reducing the sequence length to 1 / 9 of the original, thus theoretically reducing the computational cost of self-attention by about 81 times.

[0119] 2. Optimize semantic representation: Elevate the generation process to a higher semantic level of surfaces, enabling the model to learn the topology of the mesh more directly.

[0120] 3. Ensure high quality: Combining a high compression ratio (0.11) with a powerful VecSet encoder, it achieves State-of-the-Art (SOTA) reconstruction quality while significantly reducing computational overhead.

[0121] The electronic device disclosed in this embodiment includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0122] The processor may be a central processing unit (CPU) or other processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the electronic device to perform all or part of the steps of the three-dimensional mesh generation method based on facet autoregressive compression of the foregoing embodiments of this disclosure.

[0123] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0124] like Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the electronic device in the embodiment of the present disclosure. Figure 2 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0125] like Figure 2 As shown, an electronic device may include a processing unit (such as a central processing unit, graphics processing unit, etc.) that can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). The RAM also stores various programs and data required for the operation of the electronic device. The processing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0126] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow electronic devices to exchange data wirelessly or via wired communication with other devices, such as edge computing devices. Although Figure 2 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0127] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, all or part of the steps of the three-dimensional mesh generation method based on facet autoregressive compression according to embodiments of this disclosure are performed.

[0128] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0129] The computer-readable storage medium disclosed in this embodiment stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the three-dimensional mesh generation method based on facet autoregressive compression of the foregoing embodiments of this disclosure are performed.

[0130] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0131] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0132] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0133] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0134] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0135] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0136] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0137] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0138] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A three-dimensional mesh generation method based on surface element autoregressive compression, characterized in that, include: The input 3D shape representation is converted into an ordered sequence of triangle faces, where each triangle face is defined by the coordinates of its three vertices; The input 3D shape representation is encoded into a latent representation using a shape encoder; Using an autoregressive decoder, each triangle face in the ordered sequence of triangle faces is generated sequentially, conditioned on the latent representation, to generate a 3D mesh. The step of generating each triangle face in the ordered sequence of triangle faces sequentially, conditioned on the latent representation, includes: Each triangle face is treated as an independent and indivisible semantic unit, such that the length of the generated sequence of the entire mesh is equal to the number of triangle faces it contains. The input 3D shape representation is converted into an ordered sequence of triangle faces, including: Determine which point each triangular facet represents in a 3D shape; The determination of each triangular facet representing a point in the three-dimensional shape includes: Select the vertex with the smallest coordinate value among the three vertices of the triangular face as the representative point of the face; The step of using an autoregressive decoder to sequentially generate each triangle face in the ordered triangle face sequence, conditioned on the latent representation, includes: For the triangle face to be predicted, at least one triangle face that has been generated before is mapped to a corresponding face token by the face embedding layer. The face token sequence is input into the Transformer decoder, and the latent features of the current face are output through a causal self-attention mechanism and a cross-attention mechanism conditioned on the latent representation. The decoding head decodes the latent features of the current face into the coordinates of the three vertices.

2. The three-dimensional mesh generation method based on surface element autoregressive compression according to claim 1, characterized in that, The process of converting the input 3D shape representation into an ordered sequence of triangle faces also includes: Based on the spatial coordinates of the representative points of all triangle faces, all triangle faces are sorted according to a predetermined lexicographical order to form an ordered sequence of triangle faces.

3. The three-dimensional mesh generation method based on surface element autoregressive compression according to claim 1, characterized in that, The process of decoding the latent features of the current face into the coordinate values ​​of three specific vertices using a decoding head includes: The CausalMLP is used as the decoding head. When predicting multiple coordinates within a triangle face, the CausalMLP employs an in-plane autoregressive mechanism, meaning that the prediction of subsequent coordinates depends on the previously predicted prior coordinates within the same face.

4. The three-dimensional mesh generation method based on surface element autoregressive compression according to claim 1, characterized in that, The input is a three-dimensional shape representation, including point clouds, voxels, or multi-view images.

5. The three-dimensional mesh generation method based on surface element autoregressive compression according to claim 4, characterized in that, The shape encoder is a VecSet-based encoder used to convert the input point cloud into a set of latent vectors to obtain a latent representation.

6. The three-dimensional mesh generation method based on surface element autoregressive compression according to claim 4, characterized in that, Encoding the input 3D shape representation into a latent representation using a shape encoder includes: Sample k query points from the input point cloud. The query point is used as the Query, and the original complete point cloud as the Key and Value. Global geometric information is aggregated to generate initial features. ; Initial features After processing by the Transformer encoder, the latent representation C is output.

7. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the three-dimensional mesh generation method based on surface element autoregressive compression as described in any one of claims 1-6.

Images