Object recognition is becoming increasingly accurate, to the point of reducing the error rate to 6.6%.
Since 2010, the Large Scale Visual Recognition Challenge event regularly hosts the latest advances in image recognition. The last edition has been dedicated to the recognition of objects within the images and the level of precision demonstrated has doubled compared to the previous encounter, while the error rate has been reduced by half. Citing these developments the organizers have defined this year’s event as historic, emphasizing the great leap in performance.
The format of the Large Scale Visual Recognition Challenge consists of the presentation of algorithms running on high-performance computers, normally built based on GPUs, to optimize power. The different projects that are presented compete in different categories, which on this occasion ranged from the detection of objects and their location to their classification.
Have been 38 participants from 13 countries those that have competed this year, with projects that included the location and classification of large sets of images taken from various Internet sites, such as Twitter, or the modeling of biological vision systems. And the ultimate goal is for machines to be able to identify images just as the human eye does, something for which there is still a long way to go.
It is easier to correct errors year by year and improve artificial intelligence systems little by little. In this edition the results have obtained 43.9% accuracy compared to the 22.5% achieved in the previous meeting. The error rate has also improved, falling from 11.7% to 6.6%, as noted by Stanford University researcher (one of the organizers of the event, along with the University of North Carolina, Google and Facebook) Olga Russakovsky.
Most of the work this year has been based on the technology of convolutional neural networks, a concept of artificial intelligence that defined in its moderna version the French Yann LeCun in 1998. However, it has not been until now that this technique has been properly deployed thanks to cheaper computing power.
One detail to keep in mind is that participants in the event have the option to reveal the details of their algorithms or keep them as proprietary software. And in this case all the teams that won in any of the categories they chose to share knowledge with others, a good formula to drive innovation at a general level.
Images: Daniel Lee, geralt
