PaperlayerView exact revision on arXiv

Source-linked AI summary

Joint Geometrical and Statistical Alignment for Visual Domain Adaptation

Jing Zhang, Wanqing Li, Philip Ogunbona

arXiv:1705.05498v1cs.CV

TL;DR

Cross-domain visual recognition violates the shared-distribution assumption, while unsupervised adaptation lacks labeled target data. JGSA learns coupled source and target projections that jointly reduce statistical and geometrical shifts, and experiments report better performance than state-of-the-art domain adaptation methods across most evaluated datasets. Its runtime is higher than the best baseline, and action-recognition performance may not improve on test data after iterative target labeling.

  • Problem

    Visual recognition often has discrepant training and test distributions, while unsupervised domain adaptation must operate without labeled target data.

  • Method

    JGSA learns two coupled projections into respective low-dimensional subspaces while preserving target variance and source discrimination and minimizing statistical and geometrical divergence.

  • Results

    JGSA outperforms state-of-the-art domain adaptation methods on most datasets across object, digit, and action recognition tasks.

  • Takeaways & Limitations

    JGSA provides a unified alignment framework that uses both shared and domain-specific features across synthetic and real-world visual recognition tasks.

  • Takeaways & Limitations

    JGSA averages 28.97s on cross-domain object datasets, about three times the best baseline runtime, and action-task accuracy may not improve on the test set after iterative labeling.

Abstract

from arXiv · show

This paper presents a novel unsupervised domain adaptation method for cross-domain visual recognition. We propose a unified framework that reduces the shift between domains both statistically and geometrically, referred to as Joint Geometrical and Statistical Alignment (JGSA). Specifically, we learn two coupled projections that project the source domain and target domain data into low dimensional subspaces where the geometrical shift and distribution shift are reduced simultaneously. The objective function can be solved efficiently in a closed form. Extensive experiments have verified that the proposed method significantly outperforms several state-of-the-art domain adaptation methods on a synthetic dataset and three different real world cross-domain visual recognition tasks.

1. Introduction

The paper frames unsupervised domain adaptation as a practical response to distribution shifts in visual recognition, where labeled target data are unavailable. JGSA jointly reduces statistical and geometrical divergence using coupled projections and both shared and domain-specific features.

  • Motivation: Visual recognition data can shift across environments, sensors, resolutions, view angles, subjects, styles, and speeds.These factors make the same-distribution assumption of statistical learning theory unreliable.
  • Motivation: Unsupervised domain adaptation is practical but challenging because target-domain labels are unavailable and sufficient unlabeled target data are required.
  • Related approaches: Existing approaches adapt instances, feature representations, or classifiers, but classifier-based adaptation is infeasible without labeled target data.
  • Related approaches: Data-centric methods seek a shared domain-invariant space, whereas subspace-centric methods exploit domain-specific subspaces but do not explicitly align projected distributions.
  • Proposed framework: JGSA learns coupled projections that simultaneously preserve target variance and source discrimination while minimizing marginal, conditional, and geometrical divergences.The framework combines statistical alignment with a constraint keeping the two projections close.
  • Proposed framework: JGSA avoids requiring a unified transformation and was evaluated on synthetic data plus object, digit, and RGB-D action recognition tasks.The objective can also be kernelized for nonlinear shifts and solved in closed form.

2. Related Work

Related work divides domain adaptation into data-centric methods that seek common spaces and subspace-centric methods that align domain-specific subspaces. Representative methods reduce distribution discrepancy, preserve data properties, or align subspace bases using different mechanisms.

  • Data-centric approaches: Transfer Component Analysis learns shared RKHS components using Maximum Mean Discrepancy to reduce source-target sample-mean distance.
  • Data-centric approaches: Joint Distribution Adaptation extends TCA by matching both marginal and conditional distributions using pseudo-labels for target data.
  • Data-centric approaches: Transfer Joint Matching combines instance reweighting with common-subspace learning, while Scatter Component Analysis incorporates source between- and within-class scatter.
  • Subspace-centric approaches: Subspace Alignment aligns PCA-derived source and target basis vectors through a transformation matrix without assuming a unified transformation.

3. Joint Geometrical and Statistical Alignment

This section introduces the Joint Geometrical and Statistical Alignment method and presents its detailed formulation.

  • 3. Joint Geometrical and Statistical Alignment: The section presents the Joint Geometrical and Statistical Alignment method in detail.
  • 3. Joint Geometrical and Statistical Alignment: JGSA is the paper’s named method for aligning source and target domains.
  • 3. Joint Geometrical and Statistical Alignment: The section develops the method after introducing its formal presentation.

3.1. Problem Definition

The problem definition considers labeled source data and unlabeled target data drawn from different distributions. The domains share feature and label spaces, but dataset shift prevents assuming a unified transformation equalizes their distributions and conditional label relationships.

  • Data and setting: The source domain contains labeled samples, while the target domain contains unlabeled samples during training.
  • Data and setting: Source and target data lie in the same D-dimensional feature space, with ns and nt denoting their sample counts.
  • Assumptions: The domains share feature and label spaces, but dataset shift means their marginal distributions differ.
  • Assumptions: JGSA does not assume a unified transformation can simultaneously equalize marginal distributions and conditional label distributions when dataset shift is large.

3.2. Formulation

JGSA learns coupled source and target projections that jointly reduce statistical and geometrical domain divergence while preserving target variance and source discrimination. Its formulation combines distribution alignment, subspace alignment, and class-structure objectives.

  • 3.2. Formulation: JGSA finds coupled projections A and B for respective domains, exploiting both shared and domain-specific features to reduce statistical and geometrical divergence.
  • 3.2.1 Target Variance Maximization: Maximizing target-domain variance preserves target data properties and discourages projection into irrelevant dimensions.
  • 3.2.2 Source Discriminative Information Preservation: Source labels constrain the projected source representation to preserve discriminative information through within-class and between-class scatter.
  • 3.2.3 Distribution Divergence Minimization: Maximum Mean Discrepancy compares source and target sample means in the learned embeddings to minimize marginal distribution divergence.
  • 3.2.3 Distribution Divergence Minimization: Conditional distribution shift is reduced using target pseudo-labels that are iteratively refined from source-trained classifiers.
  • 3.2.3 Distribution Divergence Minimization: Unlike TCA and JDA, JGSA does not require a unified subspace because a common distributionally similar space may not exist.
  • 3.2.4 Subspace Divergence Minimization: JGSA moves source and target subspaces closer while simultaneously preserving source class information and target variance without an additional alignment matrix.
  • 3.2.5 Overall Objective Function: The overall objective balances target variance, source between-class variance, distribution shift, subspace shift, and source within-class variance using trade-off parameters.

3.3. Optimization

JGSA rewrites its coupled-projection objective as a constrained generalized eigenvalue problem. The transformation is obtained analytically from leading generalized eigenvectors.

  • 3.3. Optimization: The concatenated projection matrices [A^T B^T] are rewritten as W^T to express the objective and constraints in a joint variable.
  • 3.3. Optimization: The objective is invariant to rescaling W, motivating an equivalent rescaled formulation before constructing the Lagrange function.
  • 3.3. Optimization: The resulting stationary condition is solved by generalized eigenvalue decomposition, selecting the k leading eigenvectors to form W and recover A and B.

3.4. Kernelization Analysis

JGSA extends to nonlinear domain shifts in an RKHS by replacing original features with kernel representations. The algorithm initializes pseudo-labels, repeatedly solves the generalized eigenproblem and updates alignment statistics, then trains the adaptive classifier.

  • 3.4. Kernelization Analysis: JGSA kernelizes its projections using the Representer Theorem, representing them as P = Φ(X)A and Q = Φ(X)B in an RKHS.
  • Algorithm 1: Joint Geometrical and Statistical Alignment: The algorithm takes source and target data, source labels, and parameters λ, µ, k, T, and β as inputs.
  • Algorithm 1: Joint Geometrical and Statistical Alignment: JGSA initializes target pseudo-labels with a classifier trained on original source data before constructing scatter and distribution matrices.
  • Algorithm 1: Joint Geometrical and Statistical Alignment: Each iteration solves the generalized eigendecomposition problem and selects eigenvectors associated with the k leading eigenvalues to obtain A and B.
  • Algorithm 1: Joint Geometrical and Statistical Alignment: The transformed data produce embeddings, after which the classifier is trained on source embeddings to update target pseudo-labels and ultimately obtain the adaptive classifier.
  • 3.4. Kernelization Analysis: The kernelized formulation replaces original data matrices with feature-mapped and kernel matrices, including K, K_s, and K_t.
  • 3.4. Kernelization Analysis: The kernelized objective is solved in the same way as the original objective to compute A and B.

4. Experiments

Experiments evaluate JGSA on synthetic data and three real-world cross-domain visual recognition tasks. JGSA aligns domains effectively and generally outperforms state-of-the-art adaptation methods, while showing broad parameter ranges with satisfactory performance.

  • Synthetic Data: Baseline methods leave substantial limitations: subspace alignment can retain domain divergence, while unified transformations can mix classes or fail to satisfy all constraints.These issues appear across SA, SDA, TCA, JDA, TJM, and SCA in the synthetic comparison.
  • Synthetic Data: In the synthetic experiment, JGSA aligns source and target domains well despite a large domain shift.The synthetic data contain three RBFian classes with shifted global and class means, using two-dimensional subspaces.
  • Experimental Setup: JGSA is evaluated on synthetic data plus object, digit, and RGB-D-based action recognition tasks across multiple datasets.The real-world tasks use Office+Caltech, USPS/MNIST, and four RGB-D action datasets.
  • Results and Discussion: JGSA outperforms state-of-the-art domain adaptation methods on most real-world object, digit, and action datasets.The comparison includes primal, linear-kernel, and RBF-kernel JGSA variants alongside subspace-centric and data-centric baselines.
  • Results and Discussion: 28.97s is JGSA's average runtime on cross-domain object datasets, about three times the best baseline's runtime because it learns two mappings simultaneously.The doubled eigen-decomposition matrix size contributes to the higher runtime relative to JDA.
  • Parameter Sensitivity: Broad parameter ranges produce better results than the best baseline: β ∈[2−15, 0.5] and k ∈[20, 180].Fixing λ = 1 and µ = 1 is sufficient across all three tasks, while iteration behavior differs by task.

5. Conclusion

JGSA is a unified unsupervised domain adaptation framework that reduces domain shifts by considering both geometrical and statistical properties and exploiting shared and domain-specific features. Experiments on synthetic data and three real-world visual recognition tasks validate its effectiveness against several state-of-the-art methods.

  • JGSA jointly considers geometrical and statistical properties to reduce source-target domain shifts.
  • The framework exploits both shared and domain-specific features from the source and target domains.
  • Experiments on synthetic data and three types of real-world visual recognition tasks validate JGSA against several state-of-the-art domain adaptation methods.
Loading 1705.05498v1…