1. The document proposes a method for unsupervised visual domain adaptation using auxiliary information from the target domain. It uses partial least squares (PLS) to generate target domain subspaces incorporating subsidiary non-visual data like depth features.
2. Experiments on object recognition across the ImageNet and B3DO datasets show the method outperforms previous subspace-based approaches that only use visual information. Using auxiliary target domain data improves classification accuracy consistently.
3. The approach is an improvement as most prior work on unsupervised domain adaptation did not leverage non-visual cues available in the target domain. This opens possibilities to incorporate other multimedia signals from life logging systems.