There is no royal road to computer vision

Abstract

To endow computers with human visual capability is one of the main goals of artificial intelligence (AI) although there still is a long way to go. Taking object recognition as an example in 1980s, a main approach addressing the problem is the 3D reconstruction one, i.e., the reconstruction of 3D object from 2D images. In 1990s since the 3D reconstruction method was confronted with extreme difficulty, most researchers abandoned the attempts and turned to the 2D based approach, i.e., object recognition from 2D images directly. However, the new road is still uneven. In this talk, I will address the main principles of the new approach, its seedtime and the difficulty faced recently. When a huge amount of 2D-image data are obtained by digital cameras in object recognition (or classification), they should be transformed into an object invariant representation. In order to solve the problem, we need two key techniques, i.e., a robust detector and an object invariant describer. A number of great efforts have been made on these techniques, but so far few efficient solutions have been found. A new direction emerged to solve the problems of computer vision is that computer science may learn some things from neuron science or brain science. This talk will discuss what computer vision can learn from human visual principles and how it will be affected by the new interdisciplinary research on computer vision.