8. A. Kushal, C. Schmid, J. Ponce: Weakly Supervised Learning of Models for Multi-View Object Recognition
Today's visual object recognition systems largely focus on
fronto-parallel views of mostly upright, nearly rigid objects with
characteristic texture patterns. To overcome these limitations, we
propose a novel framework for visual object recognition with the
following characteristics: (1) model learning is seen as a weakly
supervised image segmentation task, capturing both geometric and
photometric regularities despite intra-class variations; (2) geometric
consistency is enforced locally both at the level of model parts and
at the level of individual features within the parts, making the
system more robust to viewpoint change and intra-class variability.