Answers are grounded in this paper and cite their sources.
Modeling visual data on non-Euclidean spaces, such as Grassmann and Symmetric Positive Definite (SPD) Riemannian manifolds, has shown superior advances in visual classification. To extract complementary features, recent approaches fuse multiple manifolds using deep learning architectures, achieving superior performance compared to single-manifold networks. However, they usually fuse manifold networks in a naive joint manner, which may not extract the fine-grained intrinsic features. Alternatively, metric learning is often used to enhance the discrimination of the loss function in Riemannian manifold networks. However, the random sampling scheme of batch in Riemannian manifold networks may not ensure sufficient intra-class samples for positive pairs, which may result in instability for training. In this paper, we propose to fuse two branches of hybrid manifold networks, with each fusing the Grassmann and SPD manifold networks, in a hierarchical decoupling manner. Two branches of the hybrid manifold network are individually optimized by decoupling strategies to alleviate the problem of gradient conflict between branches. Subsequently, the outputs of two branches are hierarchically concatenated to capture the fine-grained intrinsic information. Furthermore, we remove the intra-class scatter constraint in the common metric learning regularization term. Then, we exploit a dynamic positive pair construction strategy to extract sufficient intra-class samples from two branches to circumvent the instability of metric learning caused by insufficient positive pairs. Extensive experimental results on multiple visual classification tasks validate the effectiveness of the proposed method.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: