Patent Application
20210166476
The present disclosure relates to a method for processing an image, and more particularly, to a method for automatic three-dimensional (3D) image reconstruction process from real-world two-dimensional (2D) images.
The 3D reconstruction from real-world 2D images is a challenging topic in computer vision. 3D mesh representation of an object gives the ability to the viewers to look at the 3D object from any point of view. 3D mesh models can be used for many different applications such as entertainment, education, e-commerce, etc. A dense 3D mesh model estimation from a 2D real-world image is necessary for many applications to provide realistic 3D objects. Owing to its capable of modelling shape details, a dense 3D mesh is desirable for many applications since it is lightweight and capable of modelling shape details. The dense 3D mesh is beneficial in various applications. For instance, in the entertainment industry, the dense 3D mesh representation allows the user to control the viewing perspective, which can provide a more immersive and interactive visualization experience. In e-commerce, this interactive experience provides a more realistic shopping experience by visualizing an item with different viewing perspective.
To achieve this, textured 3D geometry information of an item to be displayed is necessary, which can be obtained by capturing an object using large amounts of specialized camera equipment. Even though this can produce a high-quality 3D reconstruction of an item, it is not always feasible to capture an item with an expensive camera setup. Thus, such technology is limited only to professional camera setups.
At the moment there exist multiple commercially available solutions in which all models have been created manually, textured manually and have been manually tuned with special camera setups. Therefore, generating such model can take up to multiple days, depending on its complexity, to be generated. Such models should smoothly be generated with a limited time constrained for being used in AR (Augmented Reality) solutions. Low latency is one of the main requirements for AR applications in order to provide high quality of immersive experience.
Now, an improved arrangement has been developed to reduce the above-mentioned problems. As different aspects of the invention, the invention presents a method, a server, a computer program product and a system, which are characterized in what will be presented in the independent claims.
The dependent claims disclose advantageous embodiments of the invention.
The first aspect of the invention comprises a method of converting a two-dimensional (2D) image into a three-dimensional (3D) image using an image conversion system having at least one processor and at least one memory, the method comprising: extracting a 2D RGB (Red, Green, Blue) object image attribute from a 2D object image; uploading the extracted 2D RGB object image attribute to a cloud computing service, wherein developed algorithms may be located;
calculating a 3D mesh object image attribute based on the uploaded and extracted 2D RGB object image attribute; texturing the estimated 3D mesh object from the calculated 3D mesh object image attribute and displaying the textured 3D mesh object on a consumers' display devices.
According to an embodiment, the step of the extracting a 2D RGB object image attribute further includes a segmentation algorithm using a deep neural network. According to an embodiment, the segmentation algorithm can be deployed, such as a Mask R-CNN (convolutional neural network) or other state-of-the-art segmentation algorithms.
According to an embodiment, the segmentation algorithm is performed depending on a segmentation algorithm selection.
According to an embodiment, the step of calculating a 3D mesh object image attribute further includes determining the calculated 3D mesh object image attribute, wherein the calculated 3D mesh object image attribute is compared with a predetermined threshold value to determine whether the comparison result value is greater than the predetermined threshold value.
According to an embodiment, the step of the texturing estimated 3D mesh object further includes detecting different parts of the 2D object image and mapping the detected different parts the 2D object image on a corresponding region in the textured 3D mesh object.
According to an embodiment, the display is touchable and the system is capable of receiving and using feedback from consumers to improve a 3D reconstruction quality.
A second aspect of the invention includes a server arranged to receive information about a extracted a 2D RGB object image attribute from a 2D object image; upload the extracted 2D RGB object image attribute to a cloud computing service, wherein developed algorithms may located; calculate a 3D mesh object image attribute based on the uploaded and extracted 2D RGB object image attribute; texture the estimated 3D mesh object from the calculated 3D mesh object image attribute; and a display configured to display the textured 3D mesh object. In addition to automatic object detection, the consumers have options to manually select an object using the bounding box, and the selected object can be extracted for generating 3D mesh object.
According to an embodiment, the server is arranged to perform the method of any of the embodiments above.
A third aspect of the invention includes a computer program product for converting a two-dimensional (2D) image into a three-dimensional (3D) image, where the computer program product comprises a non-transitory computer readable media encoded with a computer program which is executable in a processor, and when the computer program is executed in the processor, it is configured to perform the steps of: extracting a 2D RGB object image attribute from a 2D object image; uploading the extracted 2D RGB object image attribute to a cloud computing service, wherein developed algorithms may be located; calculating a 3D mesh object image attribute based on the uploaded and extracted 2D RGB object image attribute; texturing the estimated 3D mesh object from the calculated 3D mesh object image attribute; and displaying the textured 3D mesh object on a display device.
According to an embodiment, the server is arranged to perform the method of any of the embodiments above.
A fourth aspect of the invention includes a system arranged to convert a two-dimensional (2D) image into a three-dimensional (3D) image using an image conversion system having at least one processor and at least one memory, the system comprising: a sensor configured to extract a 2D RGB object image attribute from a 2D object image; a controller configured to upload the extracted 2D RGB object image attribute to a cloud computing service, wherein developed algorithms may be located; calculate a 3D mesh object image attribute based on the uploaded and extracted 2D RGB object image attribute; and texture the estimated 3D mesh object from the calculated 3D mesh object image attribute; and a display configured to display the textured 3D mesh object. The consumers can provide a feedback (such as bad, average, good, excellent) for the quality of the textured 3D mesh object generated by this invention, and after collecting the defined number of feedback scores, the developed neural network can be finetuned, resulting better 3D reconstruction quality in future tasks.
According to an embodiment, the system is arranged to perform the method of any of the embodiments above.
Next the invention will be described in greater detail with reference to exemplary embodiments in accordance with the accompanying drawings, in which:
Description will now be given in detail of preferred configurations of mobile terminals according to the present invention, with reference to the accompanying drawings. Hereinafter, suffixes “module” and “unit or portion” for components used herein in description are merely provided only for facilitation of preparing this specification, and thus they are not granted a specific meaning or function. Hence, it should be noticed that “module” and “unit or portion” can be used together.
In describing the present invention, if a detailed explanation for a related known function or construction is considered to unnecessarily divert from the gist of the present invention, such explanation has been omitted but would be understood by those skilled in the art. The accompanying drawings are used to help easily understood the technical idea of the present invention and it should be understood that the idea of the present invention is not limited by the accompanying drawings. This invention describes an automatic image to a 3D object (3D mesh representation) conversion approach to generate realistic 3D models. Generating a realistic look of a 3D model from a 2D input image using fine-tuned deep neural networks which will be used in the visualization of 3D objects in AR devices for e-commerce purposes, and other similar or related solutions of AR. For this purpose, this invention proposes a framework to be used for the 3D reconstruction task. The algorithm benefits from deep neural networks to estimate a dense 3D model from a given 2D real-world image and apply the texture of a given 2D real-world image to the 3D model generated by the deep neural network algorithm.
The extracted 2D RGB object image may be then uploaded on the cloud computing service unit 130, wherein the developed algorithms may be located. The cloud computing service unit 130 may include 3D objection estimation module 132, generation of texture module 134 and texturing module 136. In the 3D objection estimation module 132 and generation of texture module 134, the image to 3D mesh algorithm developed in within this invention estimates a 3D mesh object from a given 2D RGB image. In texturing module 136, this estimated 3D mesh object may be then textured using the developed texturing algorithms. As a final step, the textured 3D objects 140 may be visualized using various devices, e.g., mobile phones, tablets, PC, etc., for augmented reality applications.
In the following, the main components of the developed invention are described:
a) 2D Image to 3D Object
This invention may use graph theory to model a 3D object from the input 2D image. The model used in this task requires the integration of two modalities: 3D Geometry and 2D image. On the 3D geometry side, the algorithm builds a graph using a graph convolutional network (GCN) on the mesh model, where the mesh vertices and edges are defined as nodes and connections in a graph, respectively. A graph consists of vertices and edges, (V, E), where V={v1, v2 . . . vN} is the set of N vertices in the mesh, and E={e1, e2 . . . eN} is the set of E edges. In this model, encoding information for 3D shape is saved per vertex and the convolutional layers of the GCN enable feature exchanging across neighboring nodes and predict the 3D location for each vertex.
On the 2D image side, 2D convolutional neural network (CNN) and Visual Geometry Group (VGG)-16 like architecture, may be used to extract perceptual features from the input image. These extracted features may be then leveraged by the GCN to progressively deform a given ellipsoid mesh into the desired 3D model. Formally, GCN takes an input feature matrix, N×F, where N is the number of nodes and L is the number of features which are attached on vertices, F={f1, f2 . . . fN} where F consist of feature vectors attached on vertices. The proposed network learns to gradually deform and increase shape details in a coarse-to-fine fashion. In the graph unpooling layers increase the number of vertices to increase the capacity of handling details.
The shape details of the 3D model may be refined with the help of adversarial learning, and training using diverse set of data set. The network has been trained based on ShapeNet database, Pix3D dataset, and over thousands of samples gathered by Intelligent Computer Vision (iCV) Lab, which contains real-world images featuring diverse objects and scenes.
To constrain the property of the output shape and the deformation procedure, the present invention may define four different differentiable loss functions. In the proposed network, the Chamfer distance loss and normal loss, Earth-mover
Distance, and Laplacian regularization loss may be utilized to guarantee perceptually appealing results. Here, the Chamfer and normal losses penalize mismatched positions and normals between triangular meshes.
b) Texturing
After 3D modelling, the present invention may be conducting texturing locally by detecting different parts of the 2D image and mapping them on the corresponding region in the 3D model. Different parts of the object may be being detected with our fine-tuned DarkNet model or similar model for polygonal meshes, generating multiple texture patches. The present invention may generate texture atlases to map a given 2D texture onto the 3D model generated in the previous section. Here, each face is projected onto its associated texture image to get projection region. For each patch, the different tuned model has been adopted so that the mapping process will be as automatic as possible. Then, the algorithm adds plausible and consistent shading effects on the 3D textured model.
In general, the various embodiments of the invention may be implemented in hardware or special purpose circuits, software, logic or any combination thereof. For example, some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software which may be executed by a controller, microprocessor or other computing device, although the invention is not limited thereto. While various aspects of the invention may be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these blocks, apparatus, systems, techniques or methods described herein may be implemented in, as non-limiting examples, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof.
The embodiments of this invention may be implemented by computer software executable by a data processor of the mobile device, such as in the processor entity, or by hardware, or by a combination of software and hardware. Further in this regard, it should be noted that any blocks of the logic flow as in the Figures may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on such non-transitory physical media as memory chips, or memory blocks implemented within the processor, magnetic media such as hard disks or floppy disks, and optical media such as DVD and the data variants thereof, CD.
A person skilled in the art appreciates that any of the embodiments described above may be implemented as a combination with one or more of the other embodiments, unless it is explicitly or implicitly stated that certain embodiments are only alternatives to each other.
It is obvious to a person skilled in the art that with technological developments, the basic idea of the invention can be implemented in a variety of ways. Thus, the invention and its embodiments are not limited to the above-described examples, but they may vary within the scope of the claims.
This application claims priority of provisional U.S. application No. 62/941,902, filed on Nov. 29, 2019 the content of which is incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 62941902 | Nov 2019 | US |
