Single image 3D model retrieval based on feature fusion and view grouping

With the development of 3D models, 3D model retrieval based on a single image has received extensive attention. However, previous methods often overlook the importance of emphasizing key information in 2D images and are constrained by the limited number of 3D model training samples. To address these issues, this paper proposes using the mask of the query image to create a background-free version, which is then fused with the original query image to emphasize key information and further reduce background interference. Using 2D views to represent a 3D model, multiple features are generated by grouping the views based the weights assigned by the attention mechanism. During the training process, a query image does not correspond to a model but multiple features of a model, which addresses the problem of limited number of 3D models to some extent. For multiple features of each model, contrastive learning is used to reduce the distance between different features of the same model and minimize the distance between the image and the model at both the instance and category levels. Experimental results demonstrate that the proposed method delivered excellent performance across all three popular benchmark datasets (Pix3D, Stanford Cars, and Comp Cars).