Feature Disentangled for 3D Object Recognition

The view-based 3D object recognition method is a hot topic in the field of computer vision. In multi-view images, there are two types of features: category features and view-related features. Category features are crucial for 3D object recognition, while view-related features may introduce interference to the recognition task. Therefore, in this paper, we attempt to eliminate the interference of view-related features by decomposing multi-view images into category features and view-related features. Specifically, we utilize variational autoencoders to extract visual features. Then, the features are separeated into category features and view-related features by maximizing the mutual information between category features and different views, maximizing the mutual information between view-related features and the original image, and minimizing the mutual information between view-related features and images from other views. With the learned crucial category features, the 3D objects can be better recognized. We have conducted extensive experiments on the standard datasets ModelNet40 and ModelNet10, and the experimental results demonstrate the superiority and feasibility of the model.