GPU deformable part model for object recognition

Abstract

We consider the problem of rapidly detecting objects in static images or videos. The task consists in locating and identifying objects of interest. With the progress of affordable high computing hardware, we propose to analyse and evaluate the deformable part model on the Graphics Processing Unit. We do not take any prior assumptions on the scene and location of the objects. We provide a fast implementation and analyse the different modules of the state-of-the-art detector. Our implementation allows to accelerate both training and testing. While maintaining comparable classification performance, we report a speed-up of × 10.6 using a standard GPU card compared to a baseline implemented in C++ on a single core and × 5 compared to a multi-core OpenMP (8 threads) implementation.