Survey on fast dense video segmentation techniques

Overview & Synopsis

Semantic segmentation aims at classifying image pixels according to given categories. Deep learning approaches have proven to be very effective for this task. However, extensions to video content are more challenging, typically requiring more complex architectures, given the temporal constraints and the additional data that video introduces. At the same time, video application tend to necessitate real-time, or at least interactive performances: self-driving cars, industrial applications, or live broadcasting to name a few, imposing even stronger constraints to video methods. In recent years, considerable efforts have been made in addressing these somewhat opposing challenges. In this survey, we explore the solutions proposed to improve the quality and accuracy of video segmentation, as well as the different techniques that can be employed to improve the efficiency of such approaches, in particular in terms of inference time. Finally, we briefly describe the datasets related to the semantic video segmentation task and the challenges involved.