Text-based three-dimensional voxel model generation method

A technology for 3D voxel and model generation, which is applied in neural learning methods, biological neural network models, character and pattern recognition, etc. It can solve the problems of insufficient 3D voxel model generation and low resolution.

CN113434918AActive Publication Date: 2021-09-24BEIJING INSTITUTE OF TECHNOLOGYGY
5 Cites 4 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2021-09-24

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention discloses a text-based three-dimensional voxel model generation method, and belongs to the field of cross-modal generation from a text to a three-dimensional voxel model. The method comprises the following steps of: processing a natural language into a text vector which can be understood by a computer through a text encoder; enabling a generator to obtain the capability of generating a matched text description model through a conditional generative adversarial network framework; using discriminators for reconstructing model labels to guide the generator to train and generate a more delicate model; making the generator generate three-dimensional voxel models under different resolutions in different stages, transmitting the three-dimensional voxel models to the discriminators in different stages, and performing combined training on results, so as to enable the generator to cope with a high-resolution model generation task, and the generated model to be more exquisite; adding a local discriminator into a high-resolution model discriminator, so as to improve the high data volume and model detail processing capability of the high-resolution model discriminator, and improve the delicacy and resolution of a three-dimensional voxel model generation result. The method has the advantages of wide applicability and good generation effect.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to a method for generating a color three-dimensional voxel model based on text, and belongs to the field of cross-modal generation from text to three-dimensional voxel model. Background technique

[0002] In recent years, with the rapid development of computer technology, technologies such as computer graphics and virtual reality have become more popular in production and life, the application scenarios of 3D models have increased significantly, and people's demand for color 3D models has also increased significantly. There are a large number of application scenarios in production and life, such as 3D games, furniture design, visualization of design-oriented educational resources, etc. However, the current 3D model production work is highly dependent on model building tools such as CAD and the modeling ability of the manufacturer. This ability is not popular in society, and people are not able to mass-produce it. This work is ver...

Examples

Embodiment Construction

[0058] In order to better illustrate the purpose and advantages of the present invention, the content of the invention will be further described below in conjunction with the accompanying drawings and examples.

[0059] Such as Figure 5 As shown, a text-based three-dimensional voxel model generation method disclosed in this embodiment, the specific implementation steps are as follows:

[0060] Step 1: Input the text description into the semantic encoder to obtain the joint semantic features of the text and the 3D voxel model. The semantic encoder is implemented by the pre-trained text encoder in the semantic encoder architecture.

[0061] In order to construct a text encoder, the present invention constructs a text encoder and a three-dimensional voxel model encoder at the same time, by constraining different descriptions of the same three-dimensional voxel model to be similar after being passed through the text encoder, and the text encoding result is mapped to the model enc...