Abstract
Subspace learning is one of the most foundational tasks in computer vision with applications ranging from dimensionality reduction to data denoising. As geometric objects, subspaces have also been successfully used for efficiently representing certain types of invariant data. However, methods for subspace learning from subspace-valued data have been notably absent due to incompatibilities with standard problem formulations. To fill this void, we introduce Approximate Grassmannian Intersections (AGI), a novel geometric interpretation of subspace learning posed as finding the approximate intersection of constraint sets on a Grassmann manifold. Our approach can naturally be applied to input subspaces of varying dimension while reducing to standard subspace learning in the case of vector-valued data. Despite the nonconvexity of our problem, its globally-optimal solution can be found using a singular value decomposition. Furthermore, we also propose an efficient, general optimization approach that can incorporate additional constraints to encourage properties such as robustness. Alongside standard subspace applications, AGI also enables the novel task of transfer learning via subspace completion. We evaluate our approach on a variety of applications, demonstrating improved invariance and generalization over vector-valued alternatives.
Introduction
In computer vision, Principal Component Analysis (PCA) is the standard tool for finding low-dimensional summaries of data. However, PCA treats every data point as an isolated vector, which fails to account for inherent variations within data classes. For instance, all possible images of a face under different lighting conditions form a low-dimensional linear subspace. Standard PCA struggles to capture this structure, particularly when training data is biased—such as when all training images are lit from a single direction.
The challenge lies in the fact that the "space of all subspaces"—the Grassmannian manifold—is non-linear and non-Euclidean. Unlike vectors, subspaces cannot be simply averaged. This work introduces Approximate Grassmannian Intersections (AGI), a geometric framework designed to learn directly from "subspace-valued" data.
Methodology
AGI reinterprets subspace learning as finding a global $k$-dimensional subspace that serves as the approximate intersection of multiple input constraint sets on a Grassmannian manifold.
Key Technical Components
- Distance Metric: The framework utilizes the projection F-norm, a computationally efficient proxy for the true geodesic distance on the Grassmannian, to measure the proximity between subspaces.
- Global Optimality: Despite the non-convex nature of the Grassmannian, the authors prove that the globally optimal solution for AGI can be obtained via a simple Singular Value Decomposition (SVD) of the concatenated basis vectors of the input subspaces.
- Robustness: To handle outliers, a robust variation incorporates sparse error terms (similar to Robust PCA), solved using the Douglas-Rachford iterative projection algorithm.
- Subspace Completion: AGI enables a novel transfer learning task where a low-dimensional input (e.g., a single image, viewed as a 1D subspace) is "completed" into a higher-dimensional subspace (e.g., an illumination subspace) by leveraging the learned global model.
An elegant theoretical result of this work is that AGI is a direct generalization of PCA; when input subspaces are 1-dimensional (vectors), the AGI solution is identical to the PCA solution.
Experiments and Results
The proposed method was evaluated on synthetic datasets and the Extended Yale Face Database B, which contains images of 38 individuals under 64 different lighting conditions.
Discriminative Power and Invariance
On a synthetic dataset with heavily overlapping classes, PCA achieved only 14.6% accuracy, whereas AGI achieved 100%. This demonstrates AGI's ability to extract discriminative information from the orientation of data structures that vector-based methods miss. In the Yale B experiments, AGI's latent space produced significantly tighter clusters for individual identities compared to PCA, which was frequently confounded by lighting directions.
Subspace Completion
The subspace completion task successfully generated face images under novel lighting conditions. By transferring the illumination structure learned from one set of subjects to a new individual, AGI enabled realistic relighting from just a single source image, matching ground truth even for subjects not seen during training.
Conclusion
AGI provides a mathematically principled way to learn representations that are invariant to shared variations by treating data as subspaces rather than points. While currently limited by the assumption of linear subspaces, the geometric framework offers a robust foundation for future extensions into non-linear manifold approximations or integration into deep learning architectures as subspace-aware layers.
Summary updated