A new kind of visual-search engine has been developed to automatically scour sports footage for clips showing specific types of action and events. According to its creators, borrowing a few tricks from the field of machine translation seems to make all the difference in improving the accuracy of video search.
Despite recent advances in visual-search engines, accurate video search still remains a challenge, particularly when dealing with sports footage, says Michael Fleischman, a computer scientist at MIT. “The difference between a home run and a foul ball is often hard for a human novice to notice, and nearly impossible for a machine to recognize.”
To cope with growing video repositories, cutting-edge systems are now emerging that use automatic speech recognition (ASR) to try to improve the search accuracy by generating text transcripts. (See “More-Accurate Video Search.”)
The trouble is, search terms are often repeated out of context, says Fleischman. This is particularly the case in sport footage, such as baseball, in which commentators frequently talk about home runs and other events regardless of what is actually happening on the field.
See how the system works
To address this issue, Fleischman and Deb Roy, director of MIT’s Cognitive Machines Group, developed a system that provides a way to associate search terms with aspects of the video, and not just with what is being said as the video plays. “We collect hundreds of hours of baseball games and automatically encode all the video based on features, such as how much grass is visible and whether there is cheering in the background,” says Fleischman.
Using machine-learning algorithms, researchers analyze these video clips to identify discrete temporal “events” by extracting patterns in the different types of shots and the order in which they occur. For example, a fly ball could be described as a sequence involving a camera panning up and a camera panning down, which also occurs during a field scene and before a pitching scene.
The search system then tries to map these events to words that appear in the transcript text by looking at their probabilistic distribution. According to Fleischman, this technique is commonly used in automatic machine translation, in which words from one language are automatically mapped onto words from another, even though they may appear in completely different orders or at different frequencies. It this case, it’s a matter of translating video into audio, Fleischman says. The system tries to find the best “translation” of the events in the video into the words uttered by the announcer.
Once a new video clip is encoded using such patterns, the system looks for co-occurrences between the matched patterns and phrases. “In this way, the system is able to find correlations with events in the game, without requiring a human to explicitly design representations for any specific events,” says Fleischman.
Giving precise figures on the accuracy of the system is difficult because there is no standard for judging. Even so, trials carried out by Fleischman and Roy involving searching six baseball games for occurrences of home runs showed promise. Using just visual search alone yielded poor results, as was the case using just speech. “However, when you combine the two sources of information, we have seen results that nearly double the performance of either one on their own,” says Fleischman.
The researchers are now looking to extend this system to other sport-video archives, such as for basketball. But it shouldn’t just benefit sports fans, says Fleischman.
In theory, the system could help with other video-search processes, such as security-video analysis, says David Hogg, a professor of computer science and head of the Vision Group at Leeds University, in the United Kingdom. This system is a very novel approach, he says, and one that shows the way forward for the unsupervised learning systems that are needed to make this kind of search automatic.
Using speech and visual information together is a powerful combination for machine learning, Hogg says. “In machine learning, it is very likely to be easier the more information there is available about each situation.”
Speech can help remove ambiguities in visual data, and visual data can help disambiguate speech, says Richard Stern, a professor of electrical and computer engineering at Carnegie Mellon University, in Pittsburgh. It’s a natural marriage, he says, but one that’s just beginning to emerge.
Until recently, there has been relatively little use of ASR to aid in search, says Stern. “But this is all changing very rapidly,” he says. “Google has been recruiting speech scientists aggressively for the past several years–another indication that multimedia search is moving from the research lab to the consumer very rapidly.”