Unsupervised Object Discovery and Segmentation in Videos

Introduction: Unsupervised object discovery (UOD) is the task of finding repeating patterns and common visual concepts across an unsorted set of images without any human supervision. These concepts should describe objects, like pedestrians or cars, and stuff, like road or sky. Once discovered, this information opens several interesting applications like summarization and filtering of visual content, discovering novel or unusual visual patterns, or reducing human annotation effort. As the amount of visual data grows exponentially and human annotation becomes costly, such applications are getting more and more important. However, existing UOD approaches typically build on still images and have to rely on prior knowledge to yield accurate results. In this work, we propose a novel video-based approach, allowing also for exploiting motion information, which is a strong and physically valid indicator for foreground objects, thus, tremendously easing the task (see Figure 1).