CVPR Daily - Thursday

However, it is not always scalable or sustainable to spend time and money on acquiring annotated data. In this paper, Vu tries to address this by proposing algorithms that can learn from limited labeled data with improved robustness and accuracy. “ Let’s say you have a room, your kitchen or your bedroom, for example, and you want to understand its spatial layout, ” Vu explains. “ You take a 360- degree panorama of your entire room, and that image captures the entire holistic space of that room. Now, we want to understand where the ceiling is, where the floor is, and where the corners are so that we can reconstruct that room in an automatic computer vision kind of way. ” With only 2,000 images available, there is not much to work with here. How do you create an algorithm that can learn from this very small subset of data? Another challenge is that learning from limited labeled data in the spatial 3D reconstruction domain has not been addressed before. Previous work has come up with algorithms that given limited images can learn to 80 per cent accuracy, but they haven’t been able to address what else can be done with the unlabeled data. There is a massive amount of unlabeled data that would be too expensive to annotate. “ What we did with this work is we said, okay, give me the 2,000 images that you have available to you and give me all of the unlabeled data that you have, ” Vu explains. “ We are going to exploit and learn from this mass of unlabeled data to augment what we can do with labeled data. ” Vu’s day job is at Flyreel . Flyreel provides total property understanding for the insurance technology sector. Through its smartphone AI application, AI Assistant, the policyholder can thoroughly document everything in their home, from risks and hazards to the spatial understanding and spatial layout of rooms. The work proposed in this paper will enable the technology to provide the insurance carriers with a much more holistic and total understanding of users’ properties so that they can provide the best insurance outcomes possible. What does Vu find most exciting about his work? 11 DAILY CVPR Thursday Phi Vu Tran