Research On Key Techniques Of Content-based 3D Model Retrieval

In recent years,with the development of 3D modeling technology and the emergence of the low-cost acquisition equipment,3D model data have shown explosive growth and become a new modality of big data besides text,image,video and audio.Since the 3D model can realistically represent the spatial structure and appearance characteristic of the object,it has been widely used in the fields of intelligent manufacturing and digital entertainment.With the exponential growth of 3D model data,it is urgent to efficiently acquire and manage 3D model data.Therefore,the content-based 3D model retrieval is becoming a hot research topic in varies fields such as computer vision and artificial intelligence.Based on in-depth review and analysis of the current methods in this field,this thesis takes multi-view and multi-model data of the 3D model as our research subjects and conducts deep research on robust feature learning and accurate similarity measure.The main contents and innovations of this thesis are summarized as follows:1.We propose two novel methods View-wised Discriminative Ranking and Unsupervised Feature Learning with Graph Embedding for multi-view robust feature learning.The first method utilizes the Ranking Support Vector Machine to explore the or-der information of multiple views in the feature space and obtained the unified representation of multi-view information.The second method proposes to use the correlation among 3D models to guide the feature learning and learn the projection matric to improve the robustness of 3D model feature by using metric learning.The experimental results on common datasets show that our method outperforms the classic methods by1.2%-6.9% and the retrieval speed is 15 times faster than these methods.2.We propose Hierarchical Graph Structure Learning method for multi-view similarity measure.This method transforms the many-to-many graph matching into the single-view similarity measure with the corresponding viewpoint,which can avoid the difficulty of local subgraph structure mining and matching in many-to-many graph matching.Moreover,this thesis uses the node context information and model context information to enhance the single-view similarity measure between two 3D models.Finally,we can obtain the similarity between 3D models by fusing the multiple singleview similarities.The experimental results on common datasets show that our method outperforms the classic methods by 2.6%-6.0%.3.We propose Multi-model 3D model retrieval method for multiple type information of 3D models.The proposed method measures the visual similarity between two multi-view sets and the structure similarity between two pointsets by using two graph matching methods,respectively.Finally,we adopt the weighted fusion strategy to fuse these two similarities as final similarity between models.The experimental results on common datasets show that our method outperforms the classic methods by 1.9%-13.3%,which is suitable for multi-model retrieval.Moreover,we build a monocular image-based 3D model retrieval database and hold the first international evaluation of monocular image-based 3D model retrieval in the international SHREC 2019-3D Shape Retrieval Contest to study the different distribution between monocular image and multiple views,which promotes the research of this task.