Multiple Subspace Alignment Improves Domain Adaptation

Abstract

We present a novel unsupervised domain adaptation (DA) method for cross-domain visual recognition. Though subspace methods have found success in DA, their performance is often limited due to the assumption of approximating an entire dataset using a single low-dimensional subspace. Instead, we develop a method to effectively represent the source and target datasets via a collection of low-dimensional subspaces, and subsequently align them by exploiting the natural geometry of the space of subspaces, on the Grassmann manifold. We demonstrate the effectiveness of this approach, using empirical studies on two widely used benchmarks, with state of the art domain adaptation performance.

Unsupervised Domain Adaptation with Progressive Domain Augmentation

“…Subspace-based techniques have also been used for domain adaptation [5,33], aiming to learn domain-invariant feature representations in a lower dimensional space. Domain adaptation is conducted then by aligning within the subspace through methods such as Geodesic Flow Kernel (GFK) [10], GFS [12], or by a learned transformation matrix [5].…”

“…Domain adaptation is conducted then by aligning within the subspace through methods such as Geodesic Flow Kernel (GFK) [10], GFS [12], or by a learned transformation matrix [5]. Thopalli et al [33] extend subspace-based domain adaptation by incorporating the concept of aligning multiple subspaces per domain with the advantage of each subspace being smaller in dimensions.…”

“…We propose to adopt a multiple subspace alignment technique [33] to align the source domain with the virtual target domain. The low dimensional subspaces…”

“…X := high-error samples Xe 8: end loop of each domain can capture the fundamental information and intrinsic geometry of the domain distribution. Aligning the subspaces across domains will naturally lead to a geometric domain alignment, while multiple subspace alignment on the Grassmann manifold can be more suitable for domain alignments than using a single set of subspaces [33].…”

“…Many approaches have been developed to carry out domain adaptation; a large portion of them choose the route of exploring domain invariant structures or representations to induce aligned distribution across the source and target domains [42,6,15,13]. This particular category of approach bears many angles; domain invariant features can be learned via the minimization of divergence between the distributions of the two domains [24,30,17,4,28,38], via adversarial training [22,8,16,6], or via an auxiliary reconstruction task [9,5,33,43]. Although such methods that directly align the source and target domains have demonstrated good adaptation performance, they can encounter great challenge and cause information loss [26] when the cross-domain divergence is large.…”