Three typical situations in the subspace alignment (SA) based domain adaptation. Black denotes the source domain, and red the target. A marker denotes a specific class. The 'alignment' indicates a transformation that moves the source subspace to the target one. The left is an ideal situation, middle the situation occurring in the SA paradigm, and right the 2DSA. SA aligns two domains well but mixes instances coming from different classes (target data cannot be classified correctly), whilst 2DSA only aligns two domains moderately but preserves good margins between different classes (target data still can be separated linearly). This finding motivates us to ponder a fundamental question: to what extend is an alignment enough for classification?

Introduction

We extend the widely-used subspace alignment approach to the two-dimensional case where the matrix-form representations are employed. We propose two-dimensional subspace alignment (2DSA)—a DA approach that performs partial alignment. We find that the partial alignment mechinism is paricularly suited for the state-of-the-art convolutional feature maps. Our experiments show that 2DSA shows a good advantage to SA. In an effort to explain why 2DSA works better, we observe that the adaptation 2DSA has an effect of pushing two distributions only close but not overlapped, as shown in the above figure. This phenomenon suggests aligning two domains to be sufficiently close is enough.

Formally, we introduce two class-level local divergence measures derived from the standard H\deltaH distance used to compare the domain differences. Our analysis shows that the between-class divergence measure correlates well with the cross-domain recognition accuracy, which somewhat explains why 2DSA works well.

Furthermore, we also demonstrate a novel DA application in agriculture of classifying three types of flowering status of maize tassels (the male flower of the maize plant) and create a dataset termed MTFS-DA for the problem. The domain shifts in agriculture usually exhibit in changing illumination caused by different weather conditions, different growth status of plants because of varying soil conditions of various geographical locations, and versatile visual characteristics caused by different cultivars.

Code

The Matlab implementation of 2DSA can be found via the link shown below:

2DSA-Matlab-Release

• Google Drive
• Baidu Yun

Publication

Hao Lu, Zhiguo Cao, Yang Xiao, Yanjun Zhu, Two-dimensional subspace alignment for convolutional activations adaptation, Pattern Recognition, volume 71, pages 320-336, 2017.

Citation

@article{lu20172dsa,

  title={Two-dimensional subspace alignment for convolutional activations adaptation},

  author={Lu, Hao and Cao, Zhiguo and Xiao, Yang and Zhu, Yanjun},

  journal={Pattern Recognition},

  volume={71},

  pages={320--336},

  year={2017},

  doi={10.1016/j.patcog.2017.06.010},

  publisher={Elsevier}

}