My thesis focuses on exploiting mid-level representations for modeling objects in images. These representations provide a middle ground between bottom-up and top-down processing, are more robust than pixel-level features and can be shared across different object classes. The mid-level structures we consider are medial axes, object parts and convolutional features, which divide the thesis into three distinct parts:
In the first part we propose a machine learning approach to medial axis detection in natural images that outperforms all previous methods.
In the second part we focus on accelerating part detection using a hierarchical, coarse-to-fine approach, that achieves significant speed-ups with negligible penalty in performance.
In the third and last part we show how we can adapt convolutional features to the task of segmenting fine part structures in two diverse types of images: natural images and 2D slices from brain MRI volumes.