Human, Motion and Other Priors for Partially-Supervised Recognition. (Diverses classes d'a priori pour la reconnaissance partiellement supervisée)

This HDR manuscript presents a summary of my research activities after my PhD in 2010. It covers my work in computer vision from the postdoctoral position at Inria Paris to my present researcher position at Inria Grenoble. Understanding visual data automatically—one of the main challenges in computer vision—is having a significant impact in many practical applications, and this phenomenon can only increase with the continuous rise in digital image and video content being generated. My work presented here focuses on a selection of machine learning methods for computer vision problems. The core theme of these methods is the extraction of priors as additional cues for recognition when only partially-supervised data is available. Such partially-supervised data includes cases where only weak annotations are available, e.g., image or video labels describing the objects in a scene, instead of pixel-wise labels for segmenting objects. It also includes scenarios where data is semi-supervised, e.g., the problem of tracking objects in a video sequence when they are annotated only in the first frame. A third example of partially-supervised data is the case of incremental learning, where an existing model is updated with new training data, in the absence of the original annotations used to train the initial model. In addition to discussing approaches to handle all these scenarios, which lack full annotations, we will also demonstrate the importance of priors for a few fully-supervised recognition problems.