MACHINE PERCEPTION AND ROBOTICS GROUP

Dept. of Robotics Science and Technology,
Chubu University

This paper proposes an object category recognition method based on a bag-of-features algorithm that uses co-occurrence expressions of foreground and background information. Since bag-of-features algorithms use histograms to express features, they ignore object position information. They are considered more precise since they only code feature values in foreground regions comprised of the target categories. We investigated a method that first uses image segmentation to extract foreground regions, then codes only the feature values for those regions. We compared this method’s recognition rate to the recognition rate of the standard bag-of-features algrithm. Our experimental findings demonstrated that coding feature values from both foreground and background regions resulted in more precise recognition than coding feature values from foreground regions only. Based on these findings, we have proposed a bag-of-features algorithm that focuses on the co-occurrence of local features in the foreground and background, and uses 2+1D vector quantization histograms. Our evaluation testing showed that the proposed algorithm had a recognition rate about 3.8% better than the standard bag-of-features algorithm.