Learning to localize goal-­oriented actions with weak supervision

Overview & Synopsis

The goal of this thesis is to develop methods for automatic understanding of video content. We focus on instructional videos that demonstrate how to perform complex tasks, such as making an omelette or hanging a picture. First, we investigate learning visual models for the steps of tasks, using only a list of steps for each task, instead of costly and time consuming human annotations. Our model allows us to share the information between the tasks on the sub-step level, effectively multiplying the amount of available training data. We demonstrate the benefits of our method on a newly collected dataset of instructional videos, CrossTask. Next, we present a method for isolating task-related actions from the surrounding background, that doesn’t rely on human supervision. Finally, we learn to associate natural language instructions with the corresponding objects within the 3D scene, reconstructed from the videos.