Low-bit representation of linear classifier weights for mobile large-scale image classification

In this paper, we propose an effective method to implement a system of large-scale visual recognition where the number of classes is more than 1000 on mobile devices. Because the size of memory and storage on mobile devices such as smartphones is limited, the size of image recognition application should be as small as possible. To save the required memory of mobile visual recognition, we proposed a scalar-based classifier weight compression method before [6]. Although it is very simple and effective, it has the drawback that the performance is degraded largely in case of lower-bit representation. Then, in this paper, we propose an improved method to make 2-bit and 1-bit representation feasible, and make more comprehensive experiments including more large-scale 10k image classification with combination of the proposed improved scalar-based compression method and product quantization.