Visual recognition is inherently a large-scale problem in that humans can recognize tens of thousands of categories. One of the goals of computer vision is to endow machines with this ability. In this talk, I will discuss the challenges and opportunities toward recognition at this human scale. I will first discuss acquiring large-scale data, in particular, the construction of a large-scale visual ontology and an evaluation of state of the art recognition algorithms on more than 10,000 categories. Then I will discuss recent work on large-scale machine learning for visual recognition. I will demonstrate that, by exploiting the structures among categories, our methods significantly improve the computational efficiency and reliability of recognition.
Copyright. All Rights Reserved.