Spectral attribute learning for visual regression

A number of computer vision problems such as facial age estimation, crowd counting and pose estimation can be solved by learning regression mapping on low-level imagery features. We show that visual regression can be substantially improved by two-stage regression where imagery features are first mapped to an attribute space which explicitly models latent correlations across continuously-changing output. We propose an approach to automatically discover “spectral attributes” which avoids manual work required for defining hand-crafted attribute representations. Visual attribute regression outperforms direct visual regression and our spectral attribute visual regression achieves state-of-the-art accuracy in multiple applications.