Three-dimensional reconstruction method, apparatus and device, and storage medium

Abstract

The invention discloses a three-dimensional reconstruction method and device, equipment and a storage medium, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a to-be-processed view; and inputting the to-be-processed view into a pre-constructed three-dimensional human body reconstruction pipeline model to obtain a reconstructed three-dimensional Gaussian, the three-dimensional human body reconstruction pipeline model being obtained by combining a multi-view diffusion model and a three-dimensional human body data set, and the three-dimensional reconstruction quality can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a three-dimensional reconstruction method, device, equipment and storage medium. Background Art

[0002] Reconstructing 3D human bodies from single-view images is a fundamental challenge in 3D computer vision and computer graphics and is studied due to its wide applications in industries such as VR / AR, video games, and filmmaking.

[0003] Currently, existing 3D human body reconstruction methods produce images of poor quality.

[0004] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of this application is to provide a three-dimensional reconstruction method, device, equipment and storage medium, aiming to solve the technical problem of poor generation quality of current three-dimensional human body reconstruction methods.

[0006] To achieve the above objectives, the present application proposes a three-dimensional reconstruction method, which comprises the following steps:

[0007] Get the view to be processed;

[0008] The view to be processed is input into a pre-constructed 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian, wherein the 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

[0009] In one embodiment, the step of inputting the view to be processed into a pre-constructed three-dimensional human body reconstruction pipeline model to obtain a reconstructed three-dimensional Gaussian includes:

[0010] Inputting the view to be processed into a multi-view diffusion model trained by the three-dimensional human body dataset to obtain a multi-view image;

[0011] Perform camera registration based on the multi-view images to obtain camera pose;

[0012] Performing face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image;

[0013] A three-dimensional Gaussian reconstruction is performed based on the view to be processed, the restored multi-view image and the camera pose to obtain the reconstructed three-dimensional Gaussian.

[0014] In one embodiment, the step of performing camera registration based on the multi-view images to obtain the camera pose comprises:

[0015] The preset three-dimensional Gaussian initial parameters and the camera initial pose are subjected to rendering loss calculation and iterative joint optimization until a preset number of iterations is reached or the rendering loss is less than a preset threshold, and the camera pose is output.

[0016] In one embodiment, the step of performing face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image comprises:

[0017] Using the view to be processed as a reference image, and selecting a preset number of images from the multi-view images as source images;

[0018] Segmenting the reference image and the source image to obtain facial regions, and acquiring facial landmarks in the reference image and the source image;

[0019] Performing multi-view stereo matching on the reference image and the source image;

[0020] Selecting a three-dimensional deformable model as an initial facial mesh, iteratively optimizing parameters of the three-dimensional deformable model to obtain an optimal facial mesh;

[0021] Rendering the optimal facial mesh into the reference image to obtain a reference image depth;

[0022] Performing forward transformation projection on the facial area in the reference image, the reference image depth and the corresponding camera pose to obtain a reconstructed source perspective facial image;

[0023] The restored multi-view image is obtained based on the reconstructed source perspective facial image.

[0024] In one embodiment, the step of iteratively optimizing the parameters of the three-dimensional deformable model to obtain an optimal facial mesh comprises:

[0025] Rendering the initial facial mesh onto the two-dimensional planes of the reference image and the source image based on the parameters of the three-dimensional deformable model and the camera pose to obtain a rendered two-dimensional facial image and a rendered two-dimensional facial landmark;

[0026] Calculating an image reconstruction loss based on the rendered two-dimensional facial image and the facial region;

[0027] Calculating a landmark loss based on the rendered two-dimensional facial landmark and the facial landmark;

[0028] Constructing a composite loss function based on the image reconstruction loss, the identification loss and the introduced regularization loss;

[0029] updating parameters of the three-dimensional deformable model by minimizing the composite loss function;

[0030] Return to the execution step: based on the parameters of the three-dimensional deformable model and the camera pose, render the initial facial mesh onto the two-dimensional plane of the reference image and the source image to obtain a rendered two-dimensional facial image and rendered two-dimensional facial landmarks, until the composite loss function converges or reaches a preset number of iterations, and use the current three-dimensional deformable model as the optimal facial mesh.

[0031] In one embodiment, the step of performing three-dimensional Gaussian reconstruction based on the view to be processed, the restored multi-view image and the camera pose to obtain the reconstructed three-dimensional Gaussian includes:

[0032] Initializing a preset multi-view stereo matching network based on the camera pose;

[0033] Inputting the view to be processed and the restored multi-view image into an initialized multi-view stereo matching network to obtain a multi-view depth map;

[0034] Optimizing the multi-view depth map to obtain an optimized multi-view depth map;

[0035] A three-dimensional Gaussian reconstruction is performed based on the optimized multi-view depth map to obtain the reconstructed three-dimensional Gaussian.

[0036] In one embodiment, the step of optimizing the multi-view depth map to obtain an optimized multi-view depth map includes:

[0037] Filtering the multi-view depth map to obtain a filtered multi-view depth map;

[0038] Projecting the filtered multi-view depth map into a global three-dimensional space to obtain a three-dimensional point cloud;

[0039] Perform floating point removal and anti-aliasing operations on the three-dimensional point cloud to obtain the optimized multi-view depth map.

[0040] In addition, to achieve the above-mentioned purpose, the present application also proposes a three-dimensional reconstruction device, which includes:

[0041] A view acquisition module is used to acquire the view to be processed;

[0042] The 3D reconstruction module is used to input the view to be processed into a pre-built 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian. The 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

[0043] In addition, to achieve the above-mentioned purpose, the present application also proposes a three-dimensional reconstruction device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the three-dimensional reconstruction method described above.

[0044] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the three-dimensional reconstruction method described above are implemented.

[0045] One or more technical solutions proposed in this application have at least the following technical effects:

[0046] This application combines a multi-view diffusion model with a three-dimensional human body dataset to instill a comprehensive understanding of human body structure knowledge into the multi-view diffusion model, thereby generating realistic multi-view human body images and improving the generation quality of three-dimensional human bodies. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] Figure 1 A schematic diagram of a process flow provided for the first embodiment of the three-dimensional reconstruction method of the present application;

[0050] Figure 2 A schematic diagram of a brief process of reconstructing a three-dimensional Gaussian through a three-dimensional human body reconstruction pipeline model in the three-dimensional reconstruction method of the present application;

[0051] Figure 3 A detailed flow chart of the three-dimensional reconstruction method of the present application for reconstructing a three-dimensional Gaussian through a three-dimensional human body reconstruction pipeline model;

[0052] Figure 4 This is the model performance evaluation diagram;

[0053] Figure 5 This is a schematic diagram of the module structure of the three-dimensional reconstruction device according to an embodiment of the present application;

[0054] Figure 6 Schematic diagram of the device structure of the hardware operating environment involved in the three-dimensional reconstruction method in the embodiment of the present application.

[0055] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0056] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0057] Technical terms involved in this application:

[0058] Face geometry prior: In face recognition or face detection tasks, a strategy that uses the geometric features of the face as prior knowledge for processing.

[0059] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0060] In this embodiment, for the convenience of description, the following description is made by taking the three-dimensional reconstruction system as the execution subject.

[0061] The main solution of the embodiment of the present application is: obtain the view to be processed; input the view to be processed into a pre-built three-dimensional human body reconstruction pipeline model to obtain a reconstructed three-dimensional Gaussian, and the three-dimensional human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a three-dimensional human body data set.

[0062] Reconstructing 3D human bodies from single-view images is a fundamental challenge in 3D computer vision and computer graphics and has been widely studied due to its wide applications in industries such as VR / AR, video games, and filmmaking. Estimating 3D structures from visible regions in a single image is already an ill-posed problem, while simultaneously estimating geometry and appearance for occluded regions is even more difficult.

[0063] To address these challenges, a large number of learning-based methods have emerged to achieve single-view Figure 3 3D human reconstruction. Previous methods PIFu and PIFuHD train 3D generative models to estimate 3D neural fields through pixel-by-pixel features. However, due to the limited size of 3D human datasets, these methods often fail to generalize to unseen scenes.

[0064] SiTH uses a diffusion model controlled by camera conditions to generate a rear view based on the input forward view, and then uses the SMPL model to reconstruct the 3D human body. However, the information provided by the front and back views alone is not enough for accurate 3D human reconstruction, and the reconstructed side view of the human body is often flat. Tech and HumanRef methods use SDS to optimize the 3D human body representation. However, given the input image as a condition, these single-view diffusion models cannot guarantee the 3D geometric consistency between the generated multiple images, so the reconstructed 3D human body texture is overly smooth.

[0065] In summary, the existing single-view Figure 3 3D human body reconstruction methods face significant challenges in achieving high-fidelity view synthesis quality.

[0066] The present application provides a solution. By combining a multi-view diffusion model with a three-dimensional human body dataset, the present application instills a comprehensive understanding of human body structure knowledge into the multi-view diffusion model, thereby generating realistic multi-view human body images and improving the generation quality of three-dimensional human bodies.

[0067] It should be noted that the execution subject of this embodiment may be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a three-dimensional reconstruction device, etc., or an electronic system capable of realizing the above functions, a three-dimensional reconstruction system, etc. The following takes the three-dimensional reconstruction system as an example to illustrate this embodiment and the following embodiments.

[0068] To improve single vision Figure 3 The present application proposes to use multi-view diffusion prior to perform single-view reconstruction. Figure 3 Reconstruction of the human body.

[0069] It should be noted that multi-view images refer to multiple images of the same scene or object taken from different perspectives. The diffusion model is a generative model that generates new data samples by learning the latent representation of the data distribution. In the field of image generation, the diffusion model has been proven to be able to generate high-quality, realistic images. Multi-view diffusion prior refers to the use of multi-view images as prior information to train the diffusion model, thereby enhancing the model's ability in 3D generation and reconstruction. This prior information helps the model learn the consistency between different views, making the generated 3D content more accurate and realistic.

[0070] Single-view diffusion models generate multiple views through multiple iterations, and the views are often inconsistent. Recent multi-view diffusion models, such as SV3D and SyncDreamer, use cross-attention mechanisms to achieve multi-view information interaction during denoising, and can generate more geometrically consistent multi-view images through a single inference.

[0071] Multi-view diffusion models have also been shown to achieve remarkable results in general 3D object generation tasks. However, extending general multi-view diffusion models to 3D human reconstruction faces the challenge of insufficient 3D human priors. Since the datasets used to train these models rarely include human data and lack basic knowledge of human structure and appearance, roughly applying the priors in these models will lead to unrealistic human reconstruction.

[0072] Based on this, the present application embodiment provides a three-dimensional reconstruction method, referring to Figure 1 , Figure 1 A schematic diagram of the process flow of the first embodiment of the three-dimensional reconstruction method of the present application.

[0073] In this embodiment, the three-dimensional reconstruction method includes steps S10 to S20:

[0074] Step S10, obtaining a view to be processed;

[0075] In this embodiment, the view to be processed may be a single reference view.

[0076] Step S20, inputting the view to be processed into a pre-constructed 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian, wherein the 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

[0077] Among them, the 3D human body reconstruction pipeline model, also known as MVD-HuGaS, is obtained by combining the multi-view diffusion model with the 3D human body dataset.

[0078] It should be noted that high-quality 3D human datasets can be used to enhance the multi-view diffusion model and instill a comprehensive understanding of human structure knowledge into the multi-view diffusion model. For example, the multi-view diffusion model SV3D can be trained based on a high-quality 3D human dataset to obtain SV3D-Human.

[0079] Reference Figure 2 , Figure 2 A schematic diagram of a brief process of reconstructing a three-dimensional Gaussian through a three-dimensional human body reconstruction pipeline model in the three-dimensional reconstruction method of the present application;

[0080] By inputting a single reference image into the 3D human body reconstruction pipeline model, a reconstructed 3D Gaussian can be obtained. Rendering the reconstructed 3D Gaussian can ensure high fidelity and real-time rendering performance from any viewpoint.

[0081] This embodiment provides a 3D reconstruction method. This application combines a multi-view diffusion model with a 3D human body dataset to instill a comprehensive understanding of human body structure knowledge into the multi-view diffusion model, thereby generating realistic multi-view human body images and improving the generation quality of 3D human bodies. Figure 3 The 3D human body reconstruction pipeline model MVD-HuGaS can achieve free-view high-fidelity rendering, and its effectiveness has been verified by the SOTA (State of TheArt, current best level) performance in benchmark tests.

[0082] Extending the general multi-view diffusion model to 3D human reconstruction faces the following two challenges:

[0083] (1) Camera misalignment. Most camera-conditional diffusion models (including SV3D) cannot ensure perfect alignment with the specified camera parameters, resulting in a mismatch between the camera parameters and the generated image, which affects the quality of 3D human reconstruction.

[0084] (2) Face distortion problem. These models still have difficulty in generating accurate faces. Dimensionality reduction in latent diffusion models is necessary for denoising, but this may lead to severe loss of facial details, especially when the face occupies a small image area. The human visual system is sensitive to facial details, which further increases the difficulty of repairing facial distortion.

[0085] In order to solve the above problem, based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar content as the first embodiment can refer to the above introduction, and will not be repeated later. On this basis, step S20, inputting the view to be processed into the pre-built three-dimensional human body reconstruction pipeline model, and obtaining the reconstructed three-dimensional Gaussian includes steps S21 to S24:

[0086] Step S21, inputting the view to be processed into a multi-view diffusion model trained by the three-dimensional human body dataset to obtain a multi-view image;

[0087] Among them, a high-quality three-dimensional human body dataset is used to enhance the multi-view diffusion model SV3D, and a multi-view diffusion model SV3D-Human trained by the three-dimensional human body dataset is obtained, which is instilled with a comprehensive understanding of human body structure knowledge into the multi-view diffusion model. This adjustment can generate realistic multi-view human body images.

[0088] It should be noted that, assuming that a reference image I0 is given, the pre-trained SV3D model can be expressed as {I i |i∈{1,…,N}}.

[0089] Assuming the object is located at the center of the canonical space, the corresponding camera pose can be expressed as {yaw angle a, pitch angle e, radius r}.

[0090] The final N frames generated match the input elevation angle and radius, expressed as {e i =e0,r i =r0|i∈{1,…,N}}, and their yaw angles are uniformly distributed, which can be expressed as

[0091] While the pre-trained SV3D model can effectively generate multi-view images of static 3D objects, directly applying it to 3D human reconstruction leads to a significant drop in performance. This is partly due to the model's lack of understanding of the 3D human body, as it is mainly trained on datasets without 3D humans. The complexity of the human body, including its non-rigid motion, occlusions, and detailed features such as faces and hands, is beyond the scope of the model's pre-training.

[0092] To address this issue, this application introduces SV3D-Human, a model infused with 3D human body knowledge to enhance human perceptual reconstruction capabilities.

[0093] SV3D-Human does not use the SMPL predicted mesh to obtain additional priors, because the inaccuracy of SMPL parameter estimation will introduce errors, which will affect the subsequent reconstruction and thus limit the generalization across different human poses.

[0094] In contrast, this application fine-tunes SV3D on the high-quality 3D human body dataset 2K2K to give it a deeper understanding of the 3D human structure.

[0095] Among them, the 2K2K dataset contains more than two thousand high-fidelity 3D clothed human models, each with one million vertices and corresponding texture maps. The diversity of the dataset is very suitable for injecting human structure priors into multi-view diffusion models.

[0096] Specifically, for fine-tuning, each 2K2K model is rendered into N+1 uniformly distributed images with a fixed height and radius. During fine-tuning, VAE (Variational Autoencoder) encodes the N+1 images into latent features and adds noise to the N target features. Given the input latent features, noisy target view latent features, camera pose embeddings, and CLIP embeddings, the model learns to estimate the noise intensity under L2 loss supervision. After fine-tuning, SV3D-Human is obtained, which integrates a large number of 2D images, 3D geometry, and 3D human structure priors. By inputting a single human image, SV3D-Human can generate coherent multi-view images around the central 3D human body.

[0097] Step S22, performing camera registration based on the multi-view images to obtain camera pose;

[0098] Among them, a 3D Gaussian can be used to represent the 3D human body, and the image reconstruction loss is used to simultaneously optimize the learnable camera and the 3D Gaussian, which can effectively correct the camera pose in the generated multi-view images.

[0099] Step S23, performing face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image;

[0100] Among them, the view to be processed can be used as a reference image, a preset number of images can be selected from the multi-view images as source images, and the three-dimensional deformable model of the multi-view is used to fit the reference depth map from the reference image face. Subsequently, the depth-guided forward projection technology is used to project the reference image face onto the source image and effectively restore the multi-view image to obtain the repaired multi-view image.

[0101] Step S24, performing three-dimensional Gaussian reconstruction based on the view to be processed, the restored multi-view image and the camera pose to obtain the reconstructed three-dimensional Gaussian.

[0102] After obtaining the aligned cameras and inpainted faces, the 3D Gaussians are reused to represent and reconstruct the target 3D human body.

[0103] Reference Figure 3 , Figure 3 A detailed flow chart of the three-dimensional reconstruction method of the present application for reconstructing a three-dimensional Gaussian through a three-dimensional human body reconstruction pipeline model;

[0104] Combination Figure 3 For example, given a frontal image of a clothed person, the proposed 3D human body reconstruction pipeline model MVD-HuGaS can reconstruct a 3D human body representation that supports free viewpoint rendering.

[0105] The process proposed in this application mainly includes four stages, such as Figure 3 As shown:

[0106] Firstly, a multi-view diffusion model SV3D-Human trained by the 3D human dataset generates multi-view images according to a single reference image.

[0107] Secondly, the embodiment of the present application sets a camera registration module to recover the accurate camera pose from the generated multi-view images.

[0108] Subsequently, the embodiment of the present application sets a face restoration module to repair the distorted face area in the multi-view image.

[0109] Finally, using the registered cameras and inpainted multi-view images, MVD-HuGaS reconstructs the final 3D Gaussian and supports real-time view synthesis.

[0110] This embodiment performs camera registration based on the multi-view image to obtain the camera pose, thereby obtaining an accurate camera pose; this embodiment performs face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image, and can repair the generated distorted face by combining the face geometry prior and the reference image (input texture information).

[0111] 3D Gaussian is a recently proposed explicit scene representation that supports high-fidelity real-time rendering. It represents a 3D scene as a set of 3D Gaussians Gs = {G j |j∈{1,...,M}}, and render the image using the Splatting technique.

[0112]

[0113] Each 3D Gaussian contains a set of parameters {μ, Ω, α, F}, where μ is the mean of the 3D Gaussian distribution and Ω is the 3D covariance, which can be decomposed into a rotation matrix R and a scaling matrix S for optimization. In addition, α and F are opacity and spherical harmonic coefficients, respectively. During the rendering process, the 3D Gaussian is projected onto a 2D plane through the perspective transformation matrix W and the affine approximation Jacobian matrix J of the projection transformation. After depth sorting of the aliased 3D Gaussian, the final color of each pixel is calculated using opacity blending. During the optimization process, the parameters of the 3D Gaussian are optimized by the reconstruction loss and the structural similarity loss between the rendered image and the input image:

[0114] L = (1-λ)L1 +λ L SSIM (2)

[0115] 3D Gaussians can achieve high-quality real-time rendering, but their limitation is that they require accurate camera poses to learn scene representations.

[0116] In view of the above problems, based on the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the second embodiment can be referred to the above introduction, and will not be repeated in the following. On this basis, step S22, performing camera registration based on the multi-view image to obtain the camera pose includes step S221:

[0117] Step S221, performing rendering loss calculation and iterative joint optimization on the preset three-dimensional Gaussian initial parameters and the camera initial pose, until a preset number of iterations is reached or the rendering loss is less than a preset threshold, and outputting the camera pose.

[0118] It should be noted that SV3D-Human inherits from SV3D and cannot guarantee accurate registration of generated images and input camera conditions. Due to the fine structure of the 3D human body, camera mismatch will significantly affect the subsequent 3D reconstruction quality.

[0119] In order to solve this problem, this application designs a joint optimization strategy. This joint optimization can be expressed by formula (3):

[0120]

[0121] Here L is the loss function in the above formula, {I i |i∈{0,1,...,N}} is a multi-view image, which includes the generated multi-view image and the input single view.

[0122] The essence of this embodiment is to render the scene as a function of the three-dimensional Gaussian and camera parameter variables, and then optimize the three-dimensional Gaussian and camera parameters simultaneously through rendering loss.

[0123] As an implementation method, only the camera extrinsic parameters may be optimized. This process may initialize the camera parameters to ideal parameters, and use the three-dimensional Gaussian representation to quickly and accurately optimize these unknown camera extrinsic parameters.

[0124] This embodiment performs rendering loss calculation and iterative joint optimization on preset 3D Gaussian initial parameters and camera initial pose until a preset number of iterations is reached or the rendering loss is less than a preset threshold, thereby obtaining an accurate camera pose and improving the 3D reconstruction quality.

[0125] Despite obtaining accurate camera parameters, high-fidelity 3D human restoration still faces the problem of facial distortion. The facial region is more complex and subtle than the body, containing detailed features in a small area, and human visual perception is highly sensitive to these areas. At the same time, generating realistic faces is a considerable obstacle faced by latent diffusion models, and information may be lost during VAE encoding into latent space, which often leads to inevitable facial distortions.

[0126] In view of the above problems, based on the second embodiment of the present application, in the fourth embodiment of the present application, the same or similar contents as those in the second embodiment can be referred to the above introduction, and will not be repeated later. On this basis, step S23, based on the view to be processed, the multi-view image and the camera pose, performs face restoration to obtain the restored multi-view image, including steps S231 to S237:

[0127] Step S231, taking the view to be processed as a reference image, and selecting a preset number of images from the multi-view images as source images;

[0128] It should be noted that multi-view stereo matching (MVS) is a method for estimating depth maps from calibrated RGB images using plane scanning and feature matching. i |i=1,2,3,...}, this method can predict the depth map of the reference image. This prediction process is based on the ideal assumption of photometric consistency, that is, it is assumed that the appearance of 3D points remains consistent in each view.

[0129] However, this assumption fails to hold in the generated multi-view images, especially in the face region which is prone to distortion.

[0130] In addition, the 3D deformable model (3DMM) is a parametric representation of the face that can accurately align a preset mesh template with the facial image through an optimization process. This alignment process uses the principal component analysis (PCA) technique to simplify the prediction of facial features by reducing the dimensionality of the data. However, the 3DMM has limitations in capturing the fine appearance of real faces, and the facial mesh derived from a single image often lacks detailed 3D information.

[0131] To this end, the present application embodiment proposes a multi-view Figure 3 The method for fitting a 3D deformable model fully utilizes the advantages of multi-view stereo matching MVS and 3D deformable model 3DMM. The embodiment of the present application aims to obtain a 3D deformable model from a reference image and two generated source images {I i |i=0,1,2} predict accurate facial depth

[0132] First, the view to be processed needs to be used as a reference image, and a preset number of images are selected from the multi-view images as source images;

[0133] Step S232, segmenting the reference image and the source image to obtain facial regions, and acquiring facial landmarks in the reference image and the source image;

[0134] For example, segmenting the face region And estimate the corresponding landmarks {LM i |i=0,1,2}.

[0135] Step S233, performing multi-view stereo matching on the reference image and the source image;

[0136] Step S234, selecting a three-dimensional deformable model as an initial facial mesh, and iteratively optimizing parameters of the three-dimensional deformable model to obtain an optimal facial mesh;

[0137] After optimization, the optimal facial mesh can be obtained.

[0138] The optimal facial mesh can then be rendered into a reference view to derive a reference facial depth.

[0139] Step S235, rendering the optimal facial mesh into the reference image to obtain a reference image depth;

[0140] Step S236, performing forward transformation projection on the facial area in the reference image, the reference image depth and the corresponding camera pose to obtain a reconstructed source perspective facial image;

[0141] Given the reference face image and the corresponding depth map, the corresponding camera intrinsic parameters and relative transformation, the reconstructed source view face image can be obtained by forward transformation projection.

[0142] For reference pictures A specific pixel p in the source image The coordinates p in ′ It can be calculated as:

[0143]

[0144] Step S237: obtaining the restored multi-view image based on the reconstructed source view facial image.

[0145] This embodiment selects a three-dimensional deformable model as an initial facial mesh, iteratively optimizes the parameters of the three-dimensional deformable model to obtain an optimal facial mesh, renders the optimal facial mesh into the reference image to obtain a reference image depth, performs forward transformation projection on the facial area in the reference image, the reference image depth and the corresponding camera pose to obtain a reconstructed source perspective facial image, and obtains the repaired multi-view image based on the reconstructed source perspective facial image, thereby making full use of the advantages of multi-view stereo matching MVS and the three-dimensional deformable model 3DMM to repair and obtain an accurate multi-view image.

[0146] Based on the fourth embodiment of the present application, in the fifth embodiment of the present application, the same or similar contents as those in the fourth embodiment can be referred to the above description, and will not be described in detail later. On this basis, step S234, selecting a three-dimensional deformable model as an initial facial mesh, iteratively optimizing the parameters of the three-dimensional deformable model, and obtaining an optimal facial mesh includes steps S2341 to S2346:

[0147] Step S2341, based on the parameters of the three-dimensional deformable model and the camera pose, rendering the initial facial mesh onto the two-dimensional planes of the reference image and the source image to obtain a rendered two-dimensional facial image and a rendered two-dimensional facial landmark;

[0148] During the optimization process, the initial facial mesh is rendered onto the 2D plane of the reference and source images with aligned camera poses, and the rendered 2D facial image is obtained through PyTorch3D. and rendered 2D logo

[0149] Step S2342, calculating an image reconstruction loss based on the rendered two-dimensional facial image and the facial region;

[0150] Step S2343, calculating a landmark loss based on the rendered two-dimensional facial landmark and the facial landmark;

[0151] Step S2344, constructing a composite loss function based on the image reconstruction loss, the identification loss and the introduced regularization loss;

[0152] Among them, the constructed compliance loss function can be expressed as L 3DMM =λ img L img +λ lm L lm +λ reg L reg (4)

[0153] Where L img , L lm and L reg They represent image reconstruction loss, mark loss and regularization loss respectively. The calculation process of image reconstruction loss and mark loss is as described in step S2342 and step S2343.

[0154] Step S2345, updating the parameters of the three-dimensional deformable model by minimizing the composite loss function;

[0155] The three-dimensional deformable model 3DMM parameters (including but not limited to shape, 3D position and rotation) are optimized by minimizing the composite loss function.

[0156] Step S2346, return to the execution step: based on the parameters of the three-dimensional deformable model and the camera pose, render the initial facial mesh onto the two-dimensional plane of the reference image and the source image to obtain a rendered two-dimensional facial image and a rendered two-dimensional facial landmark, until the composite loss function converges or reaches a preset number of iterations, and use the current three-dimensional deformable model as the optimal facial mesh.

[0157] This embodiment can fully utilize the advantages of multi-view stereo matching MVS and three-dimensional deformable model 3DMM by iteratively optimizing the parameters of the three-dimensional deformable model to obtain the optimal facial mesh.

[0158] Based on the second embodiment of the present application, in the sixth embodiment of the present application, the same or similar contents as those in the second embodiment can be referred to the above description, and will not be described in detail later. On this basis, step S24, based on the view to be processed, the repaired multi-view image and the camera pose, performs a three-dimensional Gaussian reconstruction to obtain the reconstructed three-dimensional Gaussian, including steps S241 to S244:

[0159] Step S241, initializing a preset multi-view stereo matching network based on the camera pose;

[0160] After the accurate camera pose is acquired, a preset multi-view stereo matching network may be initialized based on the camera pose.

[0161] Step S242, inputting the view to be processed and the restored multi-view image into an initialized multi-view stereo matching network to obtain a multi-view depth map;

[0162] Step S243, optimizing the multi-view depth map to obtain an optimized multi-view depth map;

[0163] Among them, the optimization process can include operations such as denoising, smoothing, and filling in missing depth information.

[0164] Step S244: Perform three-dimensional Gaussian reconstruction based on the optimized multi-view depth map to obtain the reconstructed three-dimensional Gaussian.

[0165] Among them, 3D Gaussian reconstruction uses Gaussian function (a function widely used in mathematics and statistics to describe the distribution of data) to simulate and reconstruct objects in 3D space. Specifically, this method estimates the shape and position of the object surface based on the depth value of each pixel in the depth map and the spatial relationship between these points.

[0166] This embodiment optimizes the multi-view depth map to obtain an optimized multi-view depth map, and performs three-dimensional Gaussian reconstruction based on the optimized multi-view depth map to obtain a reconstructed three-dimensional Gaussian, thereby improving the accuracy of the three-dimensional model.

[0167] Based on the sixth embodiment of the present application, in the seventh embodiment of the present application, the same or similar contents as those in the sixth embodiment can be referred to the above description, and will not be described in detail later. On this basis, step S243 optimizes the multi-view depth map to obtain the optimized multi-view depth map, including steps S2431 to S2433:

[0168] Step S2431, filtering the multi-view depth map to obtain a filtered multi-view depth map;

[0169] Among them, a fast filtering stage can be used to ensure photometric consistency and geometric consistency and eliminate part of the estimated noise.

[0170] Among them, photometric consistency is to check whether the brightness or color of corresponding pixels in images of the same scene at different viewing angles is consistent. If the depth map is estimated accurately, the brightness or color of the same object surface observed from different viewing angles should be similar. By comparing the brightness or color differences of corresponding pixels in images of different viewing angles, depth values ​​that do not meet photometric consistency can be identified and filtered out.

[0171] Geometric consistency is to check whether the depth map is consistent in spatial geometry. For example, if the depth values ​​of the same object surface in the depth maps of two perspectives are very different, or the gradient directions of the depth values ​​are inconsistent, then these depth values ​​may be wrong. By checking geometric consistency, the wrong depth values ​​can be further filtered out.

[0172] Step S2432, projecting the filtered multi-view depth map into a global three-dimensional space to obtain a three-dimensional point cloud;

[0173] The filtered points are projected into a global 3D space to obtain a sparse 3D point cloud, which is used as the initialization of a 3D Gaussian.

[0174] Step S2433: performing floating point removal and anti-aliasing operations on the three-dimensional point cloud to obtain the optimized multi-view depth map.

[0175] Among them, since the multi-view image has a white background, the initially reconstructed 3D Gaussian will generate some floating points around the human body.

[0176] To this end, the embodiment of the present application adopts a random background strategy to eliminate these white floating points to ensure clean reconstruction of the human body.

[0177] Specifically, during the reconstruction process, the color or texture of the background can be randomly changed to reduce the interference of the background on the reconstruction results. In this way, the white floating points around the human body can be effectively eliminated, making the human body reconstruction cleaner.

[0178] In addition, there are some artifacts when adjusting the camera focus, which is caused by the 2D dilation kernel and the lack of 3D frequency constraints. A 2D Mip kernel and a 3D smoothing kernel are used here to reduce these artifacts.

[0179] Among them, the Mip kernel is a filter used in image processing, which can effectively reduce high-frequency noise and artifacts in the image. The 3D smoothing kernel is used to smooth point cloud data in 3D space to reduce noise and artifacts caused by inaccurate depth estimation. Through the application of these two kernels, the smoothness and accuracy of 3D point clouds can be significantly improved.

[0180] After rapid optimization, the reconstructed 3D Gaussian can support high-fidelity real-time rendering of 3D human bodies from free viewpoints.

[0181] Reference Figure 4 , Figure 4 This is a graph to evaluate the model performance.

[0182] Figure 4 It includes input images, images without the output of the multi-view diffusion model (the model is trained with a 3D human dataset), images without camera pose registration, images without restored faces, images output by the full model, and true value images.

[0183] Among them, the input image is the original input image, which serves as the starting point of the entire processing flow.

[0184] The image without the output of the multi-view diffusion model (the model is trained by the 3D human dataset) represents the state before the multi-view diffusion model trained by the 3D human dataset is applied, showing the image situation before the model intervenes. Figure 4 It can be seen that there are errors in the legs of the image that has not been output by the multi-view diffusion model.

[0185] The image without camera pose registration reflects what the output image will look like if the camera pose (i.e. the position and orientation of the camera) is not adjusted correctly during the image registration process. This helps to understand the impact of pose registration on the final result. Figure 4 It can be seen that there are errors in the head and hands of the image without the registered camera pose.

[0186] The image of the unrestored face shows the state before the face restoration step, that is, how the image would appear if no special processing or restoration was done on the faces in the image. Figure 4 It can be seen that there are errors in the head and face of the unrestored facial image.

[0187] The image output by the full model is the final image obtained after applying all processing steps (including multi-view diffusion model, camera pose registration, face restoration, etc.), reflecting the effect of the complete model processing flow. Figure 4 It can be seen that there are errors in the head and face of the image output by the full model.

[0188] As a reference, the true image represents the accurate image that should be obtained under ideal conditions or in actual measurements, and is used to evaluate the quality and accuracy of the above steps and the final output image.

[0189] In this embodiment, the fast filtering stage can significantly improve the accuracy of the depth map. This embodiment also further processes the three-dimensional point cloud to eliminate floating points, reduce artifacts, and optimize the depth map. As a result, the reconstructed three-dimensional Gaussian can support high-fidelity real-time rendering of the three-dimensional human body free viewpoint.

[0190] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the three-dimensional reconstruction method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0191] This application also provides a three-dimensional reconstruction device, please refer to Figure 5 , the three-dimensional reconstruction device comprises:

[0192] A view acquisition module 10, used to acquire a view to be processed;

[0193] The 3D reconstruction module 20 is used to input the view to be processed into a pre-constructed 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian. The 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

[0194] The three-dimensional reconstruction device provided by the present application adopts the three-dimensional reconstruction method in the above embodiment, which can solve the technical problem of poor generation quality of the current three-dimensional human body reconstruction method. Compared with the prior art, the beneficial effects of the three-dimensional reconstruction device provided by the present application are the same as the beneficial effects of the three-dimensional reconstruction method provided by the above embodiment, and the other technical features in the three-dimensional reconstruction device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0195] The present application provides a three-dimensional reconstruction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the three-dimensional reconstruction method in the above-mentioned embodiment one.

[0196] Reference below Figure 6, which shows a schematic diagram of the structure of a 3D reconstruction device suitable for implementing the embodiment of the present application. The 3D reconstruction device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The three-dimensional reconstruction device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0197] like Figure 6 As shown, the three-dimensional reconstruction device may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the three-dimensional reconstruction device are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the 3D reconstruction device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows a 3D reconstruction device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or provided instead.

[0198] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0199] The three-dimensional reconstruction device provided by the present application adopts the three-dimensional reconstruction method in the above embodiment, which can solve the technical problem of poor generation quality of the current three-dimensional human body reconstruction method. Compared with the prior art, the beneficial effects of the three-dimensional reconstruction device provided by the present application are the same as the beneficial effects of the three-dimensional reconstruction method provided by the above embodiment, and the other technical features in the three-dimensional reconstruction device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0200] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0201] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0202] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the three-dimensional reconstruction method in the above-mentioned embodiment.

[0203] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0204] The computer-readable storage medium may be included in the three-dimensional reconstruction device; or may exist independently without being assembled into the three-dimensional reconstruction device.

[0205] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the three-dimensional reconstruction device, the three-dimensional reconstruction device: obtains a view to be processed; inputs the view to be processed into a pre-constructed three-dimensional human body reconstruction pipeline model to obtain a reconstructed three-dimensional Gaussian, and the three-dimensional human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a three-dimensional human body data set.

[0206] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0207] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0208] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0209] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned three-dimensional reconstruction method, and can solve the technical problem of poor generation quality of the current three-dimensional human body reconstruction method. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the three-dimensional reconstruction method provided in the above-mentioned embodiment, and will not be repeated here.

[0210] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A three-dimensional reconstruction method, characterized in that: The method comprises the following steps: Get the view to be processed; The view to be processed is input into a pre-constructed 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian, wherein the 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

2. The method according to claim 1, characterized in that The step of inputting the to-be-processed view into a pre-constructed 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian comprises: Inputting the view to be processed into a multi-view diffusion model trained by the three-dimensional human body dataset to obtain a multi-view image; Perform camera registration based on the multi-view images to obtain camera pose; Performing face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image; A three-dimensional Gaussian reconstruction is performed based on the view to be processed, the restored multi-view image and the camera pose to obtain the reconstructed three-dimensional Gaussian.

3. The method according to claim 2, characterized in that The step of performing camera registration based on the multi-view images to obtain camera pose comprises: The preset three-dimensional Gaussian initial parameters and the camera initial pose are subjected to rendering loss calculation and iterative joint optimization until a preset number of iterations is reached or the rendering loss is less than a preset threshold, and the camera pose is output.

4. The method according to claim 2, characterized in that The step of performing face restoration based on the view to be processed, the multi-view image and the camera pose to obtain a restored multi-view image comprises: Using the view to be processed as a reference image, and selecting a preset number of images from the multi-view images as source images; Segmenting the reference image and the source image to obtain facial regions, and acquiring facial landmarks in the reference image and the source image; Performing multi-view stereo matching on the reference image and the source image; Selecting a three-dimensional deformable model as an initial facial mesh, iteratively optimizing parameters of the three-dimensional deformable model to obtain an optimal facial mesh; Rendering the optimal facial mesh into the reference image to obtain a reference image depth; Performing forward transformation projection on the facial area in the reference image, the reference image depth and the corresponding camera pose to obtain a reconstructed source perspective facial image; The restored multi-view image is obtained based on the reconstructed source perspective facial image.

5. The method according to claim 4, characterized in that The step of iteratively optimizing the parameters of the three-dimensional deformable model to obtain the optimal facial mesh comprises: Rendering the initial facial mesh onto the two-dimensional planes of the reference image and the source image based on the parameters of the three-dimensional deformable model and the camera pose to obtain a rendered two-dimensional facial image and a rendered two-dimensional facial landmark; Calculating an image reconstruction loss based on the rendered two-dimensional facial image and the facial region; Calculating a landmark loss based on the rendered two-dimensional facial landmark and the facial landmark; Constructing a composite loss function based on the image reconstruction loss, the identification loss and the introduced regularization loss; updating parameters of the three-dimensional deformable model by minimizing the composite loss function; Return to the execution step: based on the parameters of the three-dimensional deformable model and the camera pose, render the initial facial mesh onto the two-dimensional plane of the reference image and the source image to obtain a rendered two-dimensional facial image and rendered two-dimensional facial landmarks, until the composite loss function converges or reaches a preset number of iterations, and use the current three-dimensional deformable model as the optimal facial mesh.

6. The method according to claim 2, characterized in that The step of performing three-dimensional Gaussian reconstruction based on the view to be processed, the repaired multi-view image and the camera pose to obtain the reconstructed three-dimensional Gaussian comprises: Initializing a preset multi-view stereo matching network based on the camera pose; Inputting the view to be processed and the restored multi-view image into an initialized multi-view stereo matching network to obtain a multi-view depth map; Optimizing the multi-view depth map to obtain an optimized multi-view depth map; A three-dimensional Gaussian reconstruction is performed based on the optimized multi-view depth map to obtain the reconstructed three-dimensional Gaussian.

7. The method according to claim 6, characterized in that The step of optimizing the multi-view depth map to obtain an optimized multi-view depth map comprises: Filtering the multi-view depth map to obtain a filtered multi-view depth map; Projecting the filtered multi-view depth map into a global three-dimensional space to obtain a three-dimensional point cloud; Perform floating point removal and anti-aliasing operations on the three-dimensional point cloud to obtain the optimized multi-view depth map.

8. A three-dimensional reconstruction device, characterized in that: The device comprises: A view acquisition module is used to acquire the view to be processed; The 3D reconstruction module is used to input the view to be processed into a pre-built 3D human body reconstruction pipeline model to obtain a reconstructed 3D Gaussian. The 3D human body reconstruction pipeline model is obtained by combining a multi-view diffusion model with a 3D human body dataset.

9. A three-dimensional reconstruction device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the three-dimensional reconstruction method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the three-dimensional reconstruction method according to any one of claims 1 to 7 are implemented.

Images