Rapid Mobile Object Recognition Using Fisher Vector

We propose a real-time object recognition method for a smart phone, which consists of light-weight local features, Fisher Vector and linear SVM. As light local descriptors, we adopt a HOG Patch descriptor and a Color Patch descriptor, and sample them from an image densely. Then we encode them with Fisher Vector representation, which can save the number of visual words greatly. As a classifier, we use a liner SVM the computational cost of which is very low. In the experiments, we have achieved the 79.2% classification rate for the top 5 category candidates for a 100-category food dataset. It outperformed the results using a conventional bag-of-features representation with a chi-square-RBF-kernel-based SVM. Moreover, the processing time of food recognition takes only 0.065 seconds, which is four times as faster as the existing work.