ECCV 2020 Daily - Monday

2 Poster Presentation 1 DAILY M o n d a y Adam Harley is a fourth-year PhD student in the Robotics Institute at Carnegie Mellon University in Pittsburgh, USA, under the supervision of Katerina Fragkiadaki. He speaks to us ahead of his poster today (Monday), which presents a new method for unsupervised 3D tracking. This method trains a feature space that allows 3D tracking of objects over time using only multi-view data of static points in scenes . For example, learning how to track moving cars by observing parked cars from multiple viewpoints. “Tracking is one of themost fundamental problems in computer vision,” Adam tells us. “With recent methods, you need a large amount of supervision, and that means a lot of human effort in terms of gathering annotations and then training models. Here, what we’re trying to do is reduce the burden of data collection. By collecting multi-view data – which is a lot cheaper than asking humans to annotate exactly where an object has moved from frame to frame – we can still attain the outcome of a 3D tracker.” The team would like to be able to track absolutely anything . People and vehicles are a given, but as you can see in the video, there are a wide variety of other objects which occur less frequently in the data that people have invested fewer resources into annotating. With an unsupervised method of learning trackers, there is the potential to scale up to a much broader set of categories. Tracking Emerges by Looking Around Static Scenes, with Neural 3D Mapping