Semi-supervised Learning for Recognizing Objects

Research Director). The team is specialized in computer vision, in particular visual recognition. Topic: Object detection, segmentation and recognition have been long standing problems in computer vision. Many of the successful approaches for these tasks rely on the availability of sufficiently large, manually annotated data for training models on a wide variety of exemplars [4]. In the context of large-scale datasets, which are becoming increasingly common, this is a time-consuming and an expensive requirement. Here, we will focus on methods, which learn from substantially less data. In particular we will focus on learning methods, which require only a small amount of training data available in various forms. For evaluation we will use standard datasets, such as PASCAL VOC [3], YouTube-Objects [6] and ImageNet [2]. Objective: The goal of the project is to build on approaches like self-paced learning [5], semi-supervised learning [1] for the object detection, segmentation and recognition tasks. This involves learning an initial model from a small labelled training set, and then refining it based on evaluations of the remainder of the training data, which is partially labelled at best. Interesting challenges of this approach are: (i) building an efficient refinement method to process large quantities of data; (ii) defining an effective evaluation procedure; (iii) incorporating cues fro