Image and video segmentation in weakly labeled images and videos
Image and video segmentation, which aims to detect and separate the primary objects from the background in an image or video, is an important fundamental task in computer vision. There is increasing research interest in developing advanced image and video segmentation technologies as these technologies can be used for a broad range of applications like autonomous vehicles, human computer interaction and content based video coding.
Recently, deep learning based approaches have demonstrated promising performance for many computer vision tasks including image and video segmentation. However, deep learning based approaches often require a large number of labeled samples (e.g., all training samples need to be labeled with segmentation masks for the image/video segmentation task). It is often time-consuming and expensive to collect such labeled training images/videos with detailed segmentation masks.
This project will study new deep learning approaches for simultaneous object localization and segmentation in images and videos. The proposed approaches will greatly improve the state-of-the-art and extend existing works towards commercial applications. The communication of results will thus increase interest from the computer vision and machine learning communities to deepen and broaden the studies in this challenging area
In this project, we will develop novel image and video segmentation algorithms without requiring training images/videos with segmentation masks. We aim to re-establish and continue our international lead in this area by developing efficient and effective algorithms to greatly improve the state-of-the-art and extending existing work towards commercial applications.