CVPR Daily - Thursday

Phi Vu Tran is a machine learning scientist at Flyreel AI Research. His paper focuses on how to train computer vision models with less data. He speaks to us ahead of his presentation today. The layout estimation domain is concerned with learning the boundaries of ceilings and floors and the corners in a room to gain a spatial understanding of the room layout . But the domain lacks something crucial – labeled data . To date, there are only around 2,000 labeled 360-degree panorama images from this domain available. By comparison to something like SSLayout360: Semi-Supervised Indoor Layout Estimation from 360 ° Panorama 10 DAILY CVPR Thursday Presentation ImageNet , which has millions of images and crowdsources the annotation process, that is orders and orders of magnitude less. Traditionally, machine learning and deep learning models require a lot of labeled data from human annotators. More data obviously enables more accurate and robust models and better deployment in real-world applications. Phi Vu Tran with Brian Keller