Training computer vision models typically requires tedious and time-consuming manual annotation, which hinders scaling, especially for complex tasks such as full image segmentation. In this talk, I presented recent human-machine collaboration techniques from my team, where the machine assists a human in annotating the training data and training a new model. These can substantially reduce human effort and also yield more interesting interfaces to interact with. The talk explored several cases, including segmentation of individual objects, joint segmentation of all objects and background regions in an image, using speech together with mouse inputs, and annotating object classes using free-form text written by undirected annotators.
展开▼