Computer vision is ubiquitous and can be found in nearly all camera products sold nowadays. The algorithms in use often rely on state-of-the-art object detectors. While the state-of-the-art in object detection is still being pushed by more and different datasets there is still a gap in fully understanding real-world complex scenarios. Example scenarios are vehicles making an illegal u-turn, left luggage situations at airports or falling detection for the elderly. Indeed many of the scenarios could be based on heuristics on top of state-of-the-art object detectors.
At SmarterVision we are on top of the object detectors but are continuously pushing the limits, based on current research, of video understanding. In particular, we need to capture the rare scenarios without generating too many false alarms.
In this webinar, we describe some of the techniques we use in practice to support scenarios based on techniques like zero-shot learning and few-shot learning. We will describe the trend of using self-supervised learning when a lot of data is available but not annotated. I will also describe the interesting pursuit of using games like GTA and engines like Unity to create simulated data.