Deep convolutional networks have become the go-to technique for a variety of computer vision task such as image classification, object detection, segmentation, key points detection, etc. These over-parameterized models are known to be data-hungry; tens of thousand of labelled examples are typically required. Since manual annotations are expensive, learning from “weaker” annotations (e.g. only image-level category labels to localize object instances by a bounding box) become key to expand the success of deep networks to new applications.
This tutorial will provide an overview of weakly supervised learning methods in computer vision, and we will discuss the broad area of weakly supervised object recognition and its limitations of current state-of-the-art, evaluation metrics, and future ideas that will spur disruptive progress in the field of weakly supervised learning.