CCA-based transfer learning framework

We recognize cross-view actions by exploring the shared subspace from canonical correlation analysis (CCA). Conventionally, models learned in one view usually don't generalize well to others due to dissimilar distributions in feature spaces, or even different code-books (i.e., they lie in different feature spaces.) We used CCA to find the joint subspace, where actions from different cameras share the same representation, and in this subspace cross-view recognition becomes possible.

Specifically, the source training data are then projected onto this subspace, where model is learned subsequently. The label of target testing data are also predicted in this subspace.

1. Y. R. Yeh, C. H. Huang, and Y. C. Wang, “Heterogeneous Domain Adaptation and Classification by Exploiting the Correlation Subspace,” IEEE TIP, vol. 23, no. 5, pp. 2009-2018, May. 2014. (pdf, code)

2. C. H. Huang, Y. R. Yeh, and Y. C. Wang, “Recognizing Actions across Cameras by Exploring the Correlated Subspace,” ECCV 2012 the 4th VECTaR workshop, 2012. (pdf, slides)

#canonicalcorrelationanalysis #supportvectormachine #kernelmethod