[论文解读] Joint Geometrical and Statistical Alignment for Visual Domain Adaptation

Jing Zhang, Wanqing Li+1 more|arXiv (Cornell University)|May 16, 2017

Domain Adaptation and Few-Shot Learning参考文献 20被引用 64

一句话总结

JGSA 学习两个耦合投影，联合最小化源域和目标域之间的分布和几何变化，在对象、数字和 RGB-D 动作识别任务上改进无监督视觉域适应。

ABSTRACT

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.

研究动机与目标

• 为训练和测试分布不同的跨域视觉识别动机无监督域自适应。
• 提出一个统一框架，降低域之间的分布性（统计）和几何性（子空间）移位。
• 利用共享特征和域特异特征，在保留源信息判别性的同时扩大目标域方差。
• 提供闭式优化解并扩展到核化版本以处理非线性移位。

提出的方法

• 学习两个耦合投影 A（源）和 B（目标），将数据映射到低维子空间。
• Maximize target variance: maximize Tr(B^T S_t B).
• Preserve source discriminative information: maximize Tr(A^T S_b A) and minimize Tr(A^T S_w A).
• Minimize distribution shift: minimize a combined MS (MMD-based) objective incorporating marginal and conditional distributions (equations 7–9).
• Minimize subspace divergence: minimize ||A−B||_F to align subspaces geometrically.
• Solve via generalized eigenvalue decomposition of a block matrix (Eq. 19) with iterative pseudo-label refinement for target data.
• Provide kernelized extension using Representer Theorem to handle nonlinear shifts (Eq. 21).

实验结果

研究问题

• RQ1 Can unsupervised domain adaptation be effectively achieved by jointly addressing statistical distribution shift and geometric subspace shift?
• RQ2 Does coupling two domain-specific projections outperform methods that rely on a single unified subspace or purely data-centric approaches?
• RQ3 How does JGSA perform across object recognition, digit recognition, and RGB-D action recognition under domain shift?
• RQ4 Is the kernelized (nonlinear) version of JGSA beneficial for complex domain shifts?

主要发现

• JGSA outperforms several state-of-the-art domain adaptation methods on synthetic and three real-world cross-domain tasks.
• Both primal (linear) and kernelized (RBF) versions of JGSA achieve strong results on object recognition benchmarks.
• JGSA consistently improves performance over SA, SDA, TCA, JDA, TJM, SCA, OTGL, and KEMA baselines across multiple dataset pairs.
• Iterative pseudo-label refinement for the target domain helps further reduce conditional distribution shift and improve alignment.
• The method maintains competitive runtime relative to strong baselines while offering a principled joint optimization of statistical and geometric alignment.