The computer vision is ready for his next big test: see 3D. It will be a new contest next year, IMAGEnet Challenge, stimulate the development of image recognition algorithms, which is designed to help robots see the depths of each exchange.
Since 2010, scientists have been IMAGEnet database, set of object information tags they represent more than 140,000 images consisting of the image recognition algorithm by hand. Before workout, to set another set of tasks a more detailed picture on almost every computer vision system, yes.
Every year on a large-scale visual recognition IMAGEnet is trying to participate in the challenge code, you can classify those images with the least error algorithm. It was seven years ago, this is a difficult task, but now the computer vision is a big picture in the classification.
,
In 2015, the Microsoft team has created an excellent system to 95%, more than one person on the challenge of performance history. By allowing people with photo applications Google and Apple to look for their photo collection using such as food or baby. Google photos like abstract concepts, the same kind of "happiness" image.
"When we started this project, there are no industry to deal with," said Alex Berg of the University of North Carolina at Chapel Hill University, who is one of the organizers of the race. "Now, these millions of people use the product."Presentation of the real world
Therefore IMAGEnet team to say that it is time for 2018 new challenges, although the details of this tournament have not been decided, then this will be the vision of the computer has not been solved the problem: Creating can be classified in real objects, Use natural language to describe your system.
"There is very little work on the 3D scene using machine learning algorithms," Victor said at the University of Oxford. A large database will allow the robot to build a three-dimensional image of information supplement data to determine the best way to place the object and the circumstances around them to pass the training. This database will contain images of homes and other buildings inside the scene.
The existing IMAGEnet database contains collected from the Internet and then manually tagged images, but they do not understand the depth of information required for the 3D scene. Berg said that the new competition for the database can be photographed by a numerical model to simulate true depth information from the environment or 360 degrees contained. But first someone has to do these images. Due to the difficult and expensive, the dataset can be much smaller than the initial challenge.
IMAGEnet vision robot ready in a moment, said Andrew Davidson Imperial College London. He opted for the next generation of robots, support devices such as Floor Cleaning Roomba home. They know how to handle objects and manipulate the world around them, he said. "I really think you have to shape the world, this detailed 3D understanding, but what is semantic understanding there," he said.
Dai Weisen said that the new challenges will be strengthened and virtual reality. You know where the object is to help the extended reality system in the real world, such as Microsoft HoloLens described virtual objects. "It's the same opportunities," he said.
Berg does not expect much progress in the new challenges of the first two years, but an idea of what might look like success. Finally, count on the work, you can continue to find out about their surroundings, and explain what they see, and also the human person. But none of these things is more than five years old, he said.
