Accelerating convolution-based detection model on GPU

Convolution-based detection models (CDM) have achieved tremendous success in computer vision in last few years, such as deformable part-based models (DPM) and convolutional neural networks (CNN). The simplicity of these models allows for very large scale training to achieve higher robustness and recognition performance. However, the main bottleneck of those powerful state-of-the-art models is the unacceptable computational cost of the convolution in model training and evaluation, which has become a major limitation in many practical applications. In this paper, we accelerate the convolution-based detection models with the mathematic and parallel techniques. On one hand, the convolution operation in the spatial space is converted to the dot product operation in the frequency domain for less computational cost. On the other hand, the data and tasks parallelized on graphical process units (GPU) reduce the computational time further. Experimental results on the public dataset Pascal VOC demonstrate that we can use commodity GPU to speed up the whole convolution process by 2.13x to 4.31x, compared to the multithreaded implementation on CPU.