Abstract by Michael Whitney
Deep Learning Assisted GrabCut
Semantic segmentation has become a very useful tool for many applications including autonomous vehicles, robotics, satellite image processing, and much more. Most implementations of semantic segmentation use deep neural networks and require large amounts of data. However, many existing datasets only provide acceptable, not exceptional, ground truth labels for segmentation. One of the main methods to generate this data is to use a classic approach to binary segmentation called GrabCut on each instance of segmentation in an image. This is chosen due to its ease of use since the only assistance it requires is a bounding rectangle. However, in order to get good results with GrabCut, additional user input is needed. We propose utilizing deep neural networks in conjunction with GrabCut to generate more accurate and fine-tuned semantic segmentation datasets.