Research On 3D Model Sketch Retrieval Algorithm Based On Latent Feature Space Embedding

Nowadays,with the advancement of electronic hardware-related technologies and the increase in the application scenarios of three-dimensional models such as virtual reality,the demand for three-dimensional model retrieval technology is becoming more urgent.For users,how to easily and effectively retrieve the model they need is crucial.In today's mature hardware environment,hand-drawn sketches can intuitively express the user's intentions while being easy to obtain from various devices such as mobile phones and iPads.This simple and convenient method quickly became a research hotspot in the field of3 D model retrieval.However,due to the huge intra-domain differences between sketches and 3D models,the complexity of 3D models,and the subjectivity of sketches,it makes it difficult to retrieve 3D models from sketches.Therefore,this article assumes that there is a potential consistency between the sketch and the 3D model,so a potential feature space is constructed for feature embedding of both.For feature consistency,it is assumed that there are potential visual features and potential semantic features consistent between the sketch and the 3D model,and then the following research is performed:1.A three-dimensional model sketch retrieval algorithm based on latent visual feature embedding is proposed.The algorithm includes three major steps:(1)Constructing a shared data space.What about the sketch,and render the 3D model into a 2D view form at an appropriate scale.(2)Construct a visual feature embedding space.Drawing on the idea of consistent attributes in image translation,the potential visual feature distribution across modalities is constructed.(3)similar matching.According to the potential visual feature distribution obtained in the previous step,four methods are proposed for similarity measurement and experimental comparison.2.Propose a 3D model sketch retrieval algorithm based on latent semantic feature embedding.It can be divided into three parts:(1)Constructing label semantic vectors.Use the word embedding technology in natural language processing to complete the semantic embedding of various tags.(3)Cross-modal mapping features.The construction networkextracts features from the public data space and maps them to the potential visual feature space.(3)Semantic flow pattern alignment.In the feature embedding process,constructing constraints completes the mapping from visual feature space to latent semantic feature space.Based on the above research content,the standard data sets SHREC'13 and SHREC'14 were selected for experimental verification and analysis,and the final experiments proved that the above method has a certain effect on the sketch retrieval algorithm of 3D models.