Illinois researchers win Best Student Paper for research advancing computer vision

10/12/2009 Megan Kelly, Coordinated Science Lab

ECE Professor Yi Ma graduate students Shankar Rao and Hossein Mobahi and 2006 PhD ECE alumnus Allen Yang won the 2009 Best Student Paper Award (Sang Uk Lee Award) at the Asian Conference on Computer Vision.

Written by Megan Kelly, Coordinated Science Lab

Illinois students won the award for breakthroughs in image segmentation, the task of breaking up an image into regions that are perceptually meaningful to humans.
Illinois students won the award for breakthroughs in image segmentation, the task of breaking up an image into regions that are perceptually meaningful to humans.

ECE Professor Yi Ma, graduate students Shankar Rao and Hossein Mobahi, and 2006 PhD ECE alumnus Allen Yang won the 2009 Best Student Paper Award (Sang Uk Lee Award) at the Asian Conference on Computer Vision. The conference took place September 24-28 in Xi’an, China. 

The paper, “Natural image segmentation with adaptive texture and boundary encoding,” is an extension of a previous paper co-authored by Ma, Yang, former graduate student John Wright and Ma's former adviser Shankar Sastry, dean of engineering at the University of California, Berkeley. The new group analyzed and improved the previous research. While Ma was the main supervisor and oversaw the entire project, he said his students did the majority of the research.

The paper’s authors studied the problem of image segmentation, which Rao explained is the task of breaking up an image into regions that are perceptually meaningful to humans. For example, a human can easily recognize different perspectives and regions of images in front of them. In Figure 1 an individual might break this image up into regions corresponding to the sky, tree, mountains, and building.

For a computer to correctly segment an image into such regions is considered a crucial intermediate step for computers to fully understand the content of an image, what Ma calls the “holy grail of computer vision.” Ma said its success may enable important applications such as content-based image searches through the whole Internet. This ability can dramatically reduce the computational complexity of many high-level vision functions that infer or reconstruct 3-D scenes from 2-D images.

Figure 2 shows the results of the researchers’ success with computer image segmentation.  The image illustrates an algorithm of the original image, with each region marked a different color. An ideal segmentation such as this would automatically group the pixels into regions associated with the sky, mountains, tree and building.

“Our approach in the paper is to draw a connection between this problem of image segmentation and the problem of image compression, which is the task of finding a representation of an image that uses as few bits as possible, but still preserves most of the image quality,” Rao said. “We assert in the paper that the segmentation of an image that best matches with human perception very often coincides with the segmentation that requires the fewest bits in its representation. Thus to find the best segmentation of an image, we search for the segmentation that best compresses the image.”

This research may benefit many real life applications.

“There is a lot of interest in image segmentation for medical imaging, such as locating tumors in MRI images,” Mobahi said. “Segmentation can be used to produce a first level analysis of the image, based on which more complicated tasks, such as image understanding, can build upon.”

Yang added that the results demonstrated in the paper may be appreciated by investigators in several computer vision areas: “3-D urban reconstruction seen in Google Earth could also benefit from this function by using it to analyze real street photos.”

Although their method achieves state-of-the-art performance when compared with other algorithms, Mobahi said there is still a huge gap between machine performance and human performance.

“We aim to develop a mechanism to automatically learn from images manually segmented by humans,” Mobahi said. “If we can build this connection correctly, I anticipate a significant jump in performance.”

Ma also looks forward to improving image segmentation research, but is pleased with the work accomplished so far and believes the students “absolutely deserved” the 2009 Best Student Paper Award.

“Image segmentation is a tough nut to chew on simply because this is an extensively researched topic in computer vision and there are already many extremely popular, well-established, approaches,” Ma said. “My students had a new idea, believed in it, and realized it. Now they get recognized and rewarded for their vision and hard work.”


Share this story

This story was published October 12, 2009.