Parallelized two-stage object detection in cluttered RGB-D scenes

Abstract

Modern graphics hardware (GPUs) are an amazing computational resource, but only for algorithms with suitable structure. Computer vision algorithms have many characteristics in common with computer graphics algorithms, in particular, they repeat some operations, such as feature computations, at many places in the image. However, there are also more global operations, such as finding nearest neighbors in feature space, that present more of a challenge. In this thesis, we showed how a state-of-the-art object detector, based on RGB-D images, could be parallelized for use on GPUs. By using nVidia's CUDA platform we improved the running times of critical sections up to 38 times. We also built a two-stage pipeline that improves multiple object detection in cluttered scenes. The first stage aims to achieve high precision, even at the cost of lower recall, by detecting only the less occluded objects. This results in large fraction of the scene being labeled which enables the algorithm in the second stage to focus on the less visible objects that would otherwise be missed. We analyze the performance of our algorithm and lay grounds for the future work and extensions.