Computer science doctoral student Wei Liu was part of a Google team that won two of six categories in the worldwide ImageNet Large Scale Visual Recognition Challenge 2014 (ILSVRC2014).
The GoogLeNet team of Google researchers and interns including Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Drago Anguelov, Dumitru Erhan, and Andrew Rabinovich, took first place in the areas of Object Detection and Object Classification.
Liu is a fourth-year doctoral student studying computer vision with advisors Alex Berg and Tamara Berg in the UNC Department of Computer Science. As an intern with Google this summer, Liu worked on improving object detection in images by retraining a deep convolutional network programmed by Google researchers.
Google researchers had previously trained the network to recognize whole images based on their content. Liu spent his summer applying machine learning techniques to improve the network’s performance specifically for the task of detecting individual objects in an image.
Entries to the challenge were run over large collections of images, and winners were named in six categories using different evaluation metrics. In the results released Monday evening, the GoogLeNet team performed very well, placing first in two of the six categories. Liu worked on the team’s winning entry in Object Detection, which outperformed the competition in 142 of the 200 object categories involved. The entry also maintained the highest mean average precision, meaning it consistently detected objects better than other programs in the challenge. In the category of Object Classification, the GoogLeNet entry performed with only 6.66 percent error, the best of any entry in the competition.
The team’s mean average precision of 43.9 percent and classification error of 6.66 percent amount to a roughly twofold improvement from last year’s winning entries (22.5 percent and 11.7 percent, respectively) and a fourfold improvement from 2010.
The goal of the annual ImageNet Large Scale Visual Recognition Challenge is to allow research groups and, increasingly, start-ups and other companies to evaluate the accuracy of their latest algorithms for recognizing the content of images and detecting objects in photos. In addition to bragging rights for performance, seeing where various algorithms make mistakes helps researchers decide how to focus future research efforts.
ILSVRC2014 was organized by researchers at Stanford University and the University of North Carolina at Chapel Hill with outside sponsorship from Facebook and Google. The competition attracted 38 entrants from 13 different countries. Among those entrants were Microsoft Research Asia; the National University of Singapore; Adobe Systems; the University of California, Berkeley; the Oxford University; and Google.
Research in computer vision forms the basis for a variety of technologies, including image searches, video gaming, and even automatic collision detection in automobiles. Improvements developed through benchmark challenges like ILSVRC can better equip our modern-day ubiquitous computers to understand the world around them.