Meng
Tang,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Deep convolutional neural networks (CNN) are nowadays standard in computer vision. CNNs have achieved remarkable accuracy for perception tasks, e.g., classification, detection, and segmentation. However, the success largely depends on supervised training with abundant data, which are laborious to annotate. Accurate annotations such as pixel-wise segmentation and 3D reconstruction for a 2D object are expensive or even impossible to obtain.
Weakly-supervised learning is to learn from weak supervision, for example, training CNN segmentation using image-level tags. Recently, it has attracted a lot of interest since the current fully-supervised scheme does not scale or generalize well. Various cues, priors, constraints, and principles have been utilized for CNN training when there is no full-supervision. For example, a classification network is shown to be aware of object components and parts. As such, class activation map (CAM) trained with image-level tags can tell about object localization.
In this seminar, we will discuss weakly-supervised learning for object detection, semantic segmentation, and instance segmentation. We will also introduce how to learn single image 3D reconstruction based on projection-consistency.