Learning From and Dealing With Real, Rare World Data In Computer Vision

Опубликовано 8 февраля 2018, 21:41

Computer Vision has achieved tremendous progress in recent years. Primarily because of the availability of massive datasets (e.g., ImageNet or the Yahoo YFCC100m) as well as novel algorithms (e.g., deep learning or large-scale SfM pipelines). While state-of-the-art algorithms perform well when they have access to a vast amount of data, their performance falls short when data is scarce (e.g., for rare or infrequent concepts). This is highly problematic because in the real world only a few concepts are frequent and have abundant data, but most other concepts are rare and have a limited amount of information. In this talk, I will discuss methods that use rare instances to speed up robust estimations for visual geometric problems; generalize and learn rare visual concepts, and understand and transform the behavior of visual recognition systems when they face never-before-seen concepts. I will demonstrate how computational learners that can deal with rare data are important because it enables real-world visual recognition systems to pursue and complete difficult decisions originated from infrequent inputs.

See more at microsoft.com/en-us/research/v...