Computer Vision News - January 2023

11 Transparent Object Tracking their appearance also varies due to the background. That presents an even more challenging problem. “ We saw that this problem was not solved in the community, ” Alan tells us. “ Maybe a year ago, the authors presented this benchmark for evaluating tracking algorithms on transparent objects . We saw that there was a possibility to evaluate the algorithms, but there was no data to train these algorithms. That’s howwe came to this idea to solve the problem. ” The team proposes a new training dataset for tracking transparent objects and hopes to foster the development of new algorithms in that direction. The paper’s contribution is a dataset rather than a tracking method. which is already a challenging problem. ” In this problem, the tracker is initialized on an object in a single frame , the only training example it gets. It does not know what it will track beforehand. Through self- supervised learning , the algorithm has to learn the appearance of what it is supposed to track so that it can adapt to robustly localize the target in all remaining frames, even though it may change its appearance or be occluded. The development of this field depends on the datasets available for evaluation and training . Recently, a dataset for evaluation on transparent objects emerged. Transparent objects do not just change appearance when they change their position in space; as they are see-through,